Introduce Numba-based FSM utilities

[brandonwillard]

This PR introduces Numba JITed FSM utilities with 20x speed-ups over the current pure Python implementations.

It also introduces a more memory efficient "end-to-end" means of producing FSM indices. I avoided implementing it this way originally because it involves multiple iterations through a vocabulary, but, in order to address some memory-related shortcomings of the CFG indexing approaches tested in #178, this might be the better approach for now. It's a clear trade-off between processing and memory—now leaning toward processing—but, with the JIT speed-ups, it's reasonable.

Closes #226 (for now), closes #239, and should help with #192.

• Tie this into the Regex implementation.
• Map end states to tokens in the index, so that there's no need to re-walk the FSM after sampling.
  This was always how it was supposed to work, but our previous prototype didn't implement it. Since we're updating/replacing that prototype, it might be best to add it now.
• Cache computed masks.
• Investigate Numba compilation and caching.
  We need to make sure that caching works exactly as expected (i.e. only once for all the index-building code).
• Test out some multi-threading approaches to the end-to-end indexing.
• Consider using uints for the states, instead of int64.
• Consider Python packaging/Numba AOT options.

[AL-377]

will it help to speed up the "self.regex_fsm = regex_pattern.to_fsm().reduce()" in outlines 0.0.8，i found when the set the constrain field long like maxLength=1000, it takes very long in regex_fsm construction

[rlouf]

Did you try with 0.0.9?