PR #1362

Fixes #1356

Canonicalize is taking up a significant amount due to a
regex with a huge amount of character ranges (generated by
[lalrpop](https://github.com/lalrpop/lalrpop)'s lexer expanding
multiple `\w` in a token). While this could perhaps be fixed in
lalrpop I did notice the TODO in the code and after addressing this
so we automatically union and compress on each push instead of
re-canonicalizing on every push and that fixed the performance problem.

I did see the earlier attempt at this in #1051 and it seems like that
was reverted and regression tests were added so I hope that and the
existing tests are enough (I don't have a clear idea on what tests
might be missing).

PR #1308