- We don't optimize for the NO_CAPTURE case yet. If given, we can skip the
  whole tagging stuff.

- No locale support
  * Extend jrx-local.{h,c}.

- DFA optimizations:
    * Remove duplicate states
    * Join transitions to the same target state where CCLs are subset of each other.
    * Join neighboring ranges in CCLs into a single range.
    * After the DFA has been computed, some sets/vectors could be frozen.
    * More generally, we could store the DFA states in a more compressed
      format.

- The groups returned for capturing expressions aren't always
  POSIX-conform (i.e., not always what you would expect).

- Just-in-time into LLVM code.

- Add unicode matching.