Fileids and such

- add pynini dependency
- restrict parameter textequiv_level to word
  (which is the only one currently supported),
  remove default (glyph)
- add parameter for reference to rejection_weight
  (which did not exist)
- rename beam_width in FST part to pruning_weight
  (parallel to rejection_weight), add description
- replace parameter reference to lm_beam_width
  (which did not exist) to beam_width
- add FIXMEs for things that are apparently broken now
- add description of the OCR-D processor's behaviour
  to its docstring, improve README

- remove exception when calling --help or -J
- use page_from_file properly
- use correct attributes for MetadataItem

- reimplemented the FST compilation for Pynini
- plugged the training facility of the ST error model to the CLI
  (`cor-asv-train -T st`)
- removed deprecated code

Implemented CLI parameters `-c` and `-w` for `cor-ast-fst-train`, allowing for
passing a corpus and a list of words with frequencies to be included in the
lexicon.

- removed deprecated functions
- fixed docstring formatting at some places
- described the pynini dependency in README

Passing this parameter on training creates an unweighted lexicon FST. This
should be done when combining the FST model with a language model (like
`keraslm`), because a weighted lexicon is itself a (unigram) language model.

- rename: `wrapper.FSTCorrection` -> `wrapper.PageXMLProcessor`
- use `lib.FSTLatticeGenerator` instead of a tuple of FSTs

- all globals are contained in the PlaintextProcessor object - no need for
  separating `gl_config` and `model` and passing the FST model around as a
  tuple of transducers

- use the current library (FST-based decoding) for generating the hypotheses
  graph
- update to `ocrd` v1.0.0b5
- refactoring

- remove the deprecated CLI parameters:
  - `apply_lm` - not used since a long time
  - `num_results` - replaced with `beam_width` after the switch to pynini
- removed passing some unnecessary parameters to
  `scripts.process.prepare_model()`

(no longer compatible with the current CLI)

NLTK was only used for computing character n-grams from strings.

The related code was commented out since a long time ago anyway. Also some
neighboring commented-out code was removed.

Also removed `helper.create_dict()` (doesn't use HFST, but was obsolete
anyway).

Furthermore, removed `scripts.process.prepare_composition()` (was no longer in
use, forgot to remove it with b5b1fd67).

The code there is HFST-dependent and so obsolete that it is no longer relevant
for further development.

The CLIs were no longer used, since `scripts.train` is used for training. They
were becoming increasingly deprecated (especially after switching away from
HFST).

The Cython extension for computing the FST composition is no longer needed.

The version number was increased to 0.2.0 as this is a quite important change.

The back-end for processing FSTs was changed from HFST to Pynini. The
functionality implemented so far is:
- lexicon training
- simple error model training
- processing plain text
  - window recombination using `pynini.replace()`

Further related changes:
- as Pynini does not support the `n_best()` method, beam search will be used
  instead - the hypotheses are pruned to those within `beam_width` weight
  from the best one *after each composition*, i.e. first after the composition
  with the error model and then once again after the composition with the
  lexicon (in order to keep a manageable size of the hypotheses FST);
  currently, `beam_size` is hardcoded to `5`, but it should be made a
  parameter; lower values allow for faster execution times, but may miss some
  corrections
- removed the parameter `frequency_class` from
  `lib.error_simp.transducer_from_list()` (never used)
- the behavior of `rejection_weight` was implemented to mimic the one in the
  Cython extension - i.e. the rejection weight of a word is
  `rejection_weight*(len(word)+2)`. The `+2` originally comes from the "flag"
  transitions, but turned out to be useful by preventing the rejection of short
  words.
- added a test suite (to be extended later)

No longer required:
- the Cython extension
- passing temporary files between the Python and the C++ part
- the HFST dependency (except for `error_st`, which is currently incompatible
  with the rest)

Remaining issues:
- switch the ST error model implementation to use pynini
- code cleaning: remove unused dependencies and deprecated code (esp. the
  HFST-related parts)
- restore some functionality that was temporary removed to simplify the
  transition
  - special rules for digits and umlauts in the lexicon
  - compounds in the lexicon
- make `beam_width` a free parameter
- unit tests

- the functionality of finding the shortest path in the lattice was moved to
  `lib.sliding_window.lattice_shortest_path()`
- also removed some deprecated code from `scripts.process.correct_string()`

- rename `sliding_window_no_flags` -> `sliding_window`
- move the old `sliding_window` module (with flags) to `__DEPRECATED__`
- remove the flag-related code from `error_simp` (do not add flags to the error
  transducer on training)

- obsolete and no longer needed
- should be replaced with real unit tests