Reynir

* Added binary wheels for PyPy; adopted `manylinux2010` wheel format
* Add Python 3.9 to list of supported and tested versions
* Better attachment of prepositions to infinitive verbs (*nafnháttarsagnir*)
* Grammar enhancements and simplifications
* Type annotations added

* Parsing of very long sentences (>80 tokens) is now faster and much less memory-intensive. This is due to significant improvements in the parse forest reduction algorithm. It no longer needs to unpack shared packed parse forest (SPPF) nodes in order to be able to assign different scores to preposition phrases depending on their verb context. Instead, granular memoization is used when traversing the SPPF to calculate subtree scores.
* Problems with installing under CPython on macOS fixed by adjusting C++ compiler flags.
* Documented how to use the `CFLAGS` environment variable on to install under PyPy on macOS.

* Old test corpora deleted
* README updated
* S inside CP-REL fixed

* Switched from GNU GPLv3 to MIT license, in accordance with the requirements of the Icelandic Government's 5-Year Language Technology Programme
* Python 3.6 is now a requirement; Python 3.5 is no longer supported
* Renamed repository from `ReynirPackage` to `GreynirPackage`
* Many enhancements to the context-free grammar
* Type annotations added
* Added circuit breaker to abandon the parsing of very complex and time-consuming sentences, via a `ParseError` exception
* Documentation additions and formatting improvements
* Added `force_number` parameter to `NounPhrase` class
* Added `error` attribute to the `_Sentence` class
* Many other smaller enhancements

* Lots of type annotations added
* Adapted to Tokenizer 2.1.0 which no longer splits a final period off an abbreviation that ends a sentence
* Significant additions to vocabulary; minor additions to grammar
* Added ability to dump and load `_Sentence` instances, including tokens, to and from JSON

* Added test for foreign sentences, i.e. sentences where less than 60% of the words are found in BÍN (or resolved by the composite word algorithm). Such sentences are now marked by an error code instead of being parsed.
* More flexibility in specifying the maximum length of sentences to be parsed. The (overridable) default is now a maximum of 90 tokens in a sentence.
* Added and enhanced functions for generating lemmas from parsed sentences. New properties include `_Sentence.lemmas_and_cats` and `_Sentence.categories`, and the method `Greynir.parse_tokens()` has been added.
* Grammar and vocabulary improvements.
* Better handling of abbreviations and multi-word prepositions.
* Improved type annotations in the main Greynir API.
* Added tests.
* New features documented and examples added.