sliding_window_no_flags.process_window_with_openfst() now works properly

- much cleaner and smaller code
- without flag diacritics and state merging
- the windows are recombined by adding transitions between consecutive
  windows (see the PDF documentation for details)
- test_sliding_window_no_flags.py is a temporary testing script (intended to be
  removed after the module is integrated into main processing)
- known issues:
  - process_window_with_openfst() doesn't work - the composition returns a
    transducer accepting garbage paths
  - process_window_with_hfst() is very slow
  - not yet integrated into process_test_data.py

(duplicate of helper.save_transducer())

- hfst/ -> ./
- cython/ -> extensions/
- currently unused files moved to __DEPRECATED/
- added .gitignore

- hfst/
  - ASSE data removed
- open-fst/
  - transducers removed (most are duplicates of the ones stored in
    cor-asv-fst-models; others, like the ASSE lexicon, can be trained if
    needed; shouldn't clutter the repo)
  - the report about evaluation experiments moved to doc-fst/evaluation-openfst
  - gesamt_dta_spaces.syms removed

Changes in get_precision_recall():
- Refactoring: separating the funtionality of merging alignments from scoring.
- Changed the definition of true/false and positive/negative. The characters
  that are originally wrong and wrongly corrected are now false positives
  (previously: false negatives). This also changes the evaluation results quite
  significantly!
- Bugfix: consider only one-to-one or one-to-zero alignments, but not
  one-to-many. This also changes the results.
- Code cleaning.

`process` was renamed to `correct_string` in 2fc4eb20, but not in line 235.
This commit fixes it.

- command-line argument parsing moved into a separate function
- computing each evaluation metric over all lines moved into a separate
  function, so that the logic of main() reduces to a simple three-way `if`
- increased the spacing between top-level declarations to
  two blank lines (PEP 8)
- added the -G parameter for providing the ground truth suffix

Isolated some activities into subfuctions for better structuring.

Changed the logging level of showing input/output strings from "info" to
"debug".

- grouped the globals into two dictionaries: `gl_config` and `model`
- renamed `process` to `correct_string`
- renamed `load_model` to `build_model` (does other things apart from loading)
- isolated some functionality from the `main` function
  - parallel processing of strings -> `parallel_process`
  - printing results -> `print_results`
  - building transducer composition, flag encoder and loading LM transducers ->
    `build_model`
- minimized the availability of globals to increase readability and avoid bugs
  - globals are only visible in `main()` and `correct_string()`, but not in any
    subfunctions that `main()` calls
  - instead of passing `args` (the argument parser) as a global, the dictionary
    `gl_config` is used, which contains only the values used by
    `correct_string()`

The model-building functions in sliding_window.py (load_transducers_*())
previously contained three kinds of functionalities:
- loading transducers
- variant-specific combining of transducers to a single token acceptor
- variant-independent functionality, which is copy-pasted in all three
  functions (adding flags, converting a single token acceptor to a window etc.)

This commit isolates the variant-independent functionality into smaller
functions build_single_token_acceptor_*(), combines the variant-independent
parts for all three variants in the function build_model() and puts the loading
of transducers outside of the `sliding_window` module.

Furthermore:
- renamed process_test_data.load_transducers() to load_model

- isolated loading transducers into a separate function
- isolated preparing the composition of lexicon and model into a separate
  function
- moved process() before main()
- cleaned up commented-out code, old file names etc.

- divided the body into smaller functions
- fixed (in line 140) a bug causing types that are identical in capitalized and
  uncapitalized form (like '—') to be counted double

- refactored the reading of training data

- functionalities isolated into separated functions:
  - creating a single error transducer
  - combining error transducers
- fixed a bug causing only context = 3 to be considered (line 467 pre-commit,
  previously line 41 in error_transducer_complete.py)
- simplified transducer creation

Merged `error_transducer_complete.py` into `error_transducer.py`, so that one
module is responsible for training an error model.

- isolated parse_arguments() as a separate function
- added a -G parameter (gt_suffix) instead of a fixed suffix
- removed some unnecessary comments

process_dta_data.py was merged into lexicon_transducer.py so that only a single
module is responsible for building the lexicon.

The lexica are no more saved as plaintext. This information can be easily
obtained with `hfst-fst2strings -w`.

helper.py:
- the logarithm of frequencies is computed in the normalizing function, rather
  than during writing to file

- Model files are moved to a separate repository ('cor-asv-fst-models').
  As a temporary solution, the directory 'hfst/fst' has to be linked to the
  location of the model repository so that the hard-coded paths to model files
  work.
- Changed the hard-coded model file names in process_test_data.py to match the
  names of files created by the training scripts.

- using os.path.join() instead of string concatenation
- removed (useless) trailing slashes from directory names

Minor changes:
- rename: x -> filename in helper.generate_content()

The character U+0364 (combining latin small letter e):
- is invisible in some terminal fonts,
- breaks syntax highlighting in Vim.

- setup_spacy() and parse_arguments() as separate functions
- more readable formatting

- Isolated parsing command-line arguments as a separate function.
- More readable formatting.

- Refactoring: structured the body of merge_states() into smaller subfunctions.
- Crashes were caused by deleting transitions from the transducer while
  iterating over it. In the updated version, the transitions to delete are
  first identified and then deleted.

reason: easier debugging

amend f25d826d7: in processor, when splitting input into symbols and flags, fix a new crash with decomposed/combining characters

replace slow and memory/stack-devouring alignment.sequencealigner by fast difflib.SequenceMatcher in evaluation too

allow training error transducer on CSV file instead of directory, plus Python 2 compatibility and some Pylint cosmetics

with REJECTION_WEIGHT set to -1, behave differently:
- pyComposition (if enabled) does no backoff_result()
- union with input (and input weight mapping) is instead
  done in compose_and_search afterwards, but with a long
  vector of sensible thresholds to each apply alternatively
  on the same window result (i.e. without having to rerun
  everything)
- so all results up the call graph now have to be vectorized
  as well: compose_and_search, create_result_transducer,
  window_size_1_2, main/process;
  in the normal case, there is only 1 value in that vector