Enhanced spell-corruption-aware classifier/completer/encoder training

[josephbirkner]

Important:

• Train hierarchically more frequent tokens more often according to the training grammar

Try this:

• Corrupt with phonetic substitutes (https://sqlite.org/spellfix1.html, https://service.projectplace.com/#project/411317038/documents/775109662/780603518)
• Keyboard layout distance-based substitutions
• Such substitutions must also be integrated into the DAWG via DAWG.compile_replaces().

Alternatively, try this:

• Train with deletions only (like symspell)