Analiticcl

New:
* Context rules and tagging (https://github.com/knaw-huc/golden-agents-htr#7): allows specifying regular-expression like patterns to match entities spanning multiple 'words'
* Allow choosing unicode codepoints for offsets instead of UTF-8 byte offsets (15)

Bugfixes:
* use lowest frequency of either variant or target when using variant lists (https://github.com/knaw-huc/golden-agents-htr#15)
* Allow out-of-vocabulary words in LM; not everything that's in the lexicons has to necessarily also in the LM

Important bugfix release:

* Fixed use of frequency information in score() function
* Fixed parsing of DistanceThreshold (edit distance threshold, anagram distance threshold), when it consist of a relative and absolute component.
* Better parameter validation in Python binding
* More verbose feedback on chosen parameters
* Fixed version information

* fixed auto-detection of frequency information in parsing variant lists
* fix for the python wheel building

Minor bugfix release: fixes an issue with invalid JSON serialisation 13

Major development updates:

* Initial implementation on finding matches in running text (error detection); search mode 2
* Support for Language Models to consider context
* Support for n-grams; decoding using Finite State Transducers
* Strict separation between lexicon and language model
* Still experimental
* Removed frequency from score component and added it as a separate score
* Added frequency-ranking as an opt-in feature now; explicitly propagate frequency score and distance score separately to the output
* Removed lexicon weights
* Made distance score computations relative to input length
* Changed default weights so levenshtein-damarau carries most weight
* Implemented a Python binding (1)
* Fixed insertions after deletion (6), removed premature bound-check optimisations
* Implemented a learning mode that collects variants for a given lexicon, either in running text or matched against another test lexicon
* Implemented a cut-off threshold
* Allow frequency information in variant lists
* Adhere strict to lexiconc/variantlist loading order as specified on command line
* Return all matching lexicons for matching rather than just one (in case an entry exists in multiple lexicons)
* More debug levels
* Anagram/edit distance can now be set to an absolute value or a ratio (relative to input length)
* Significant documentation updates

* This release replaces the underlying big integer library with ibig 0.3.2, which leads to a significant performance increase due to less heap allocations.
* Implemented explicit variant ingestion and matching (but still requires proper testing)
* fixed benchmarks
* allow some escape sequences in alphabet files