Incremental reparsing

[eddyb]

Allow supplying one (or more?) range of changes to a past input, and reuse the memoizer and SPPF.

Maybe we can use a non-contiguous range space? That way we instead parse the concatenation of "unchanged prefix", "changed range", "unchanged suffix", and reuse the prefix/suffix subgraphs verbatim, instead of having to relabel every node we keep with its new range.

Aside from the information we keep today, we'd probably need to track the "union of all accessed input ranges" for a given Call, as some of the attempted branches that failed may succeed later.

Also see tree-sitter from #5.

[eddyb]

One idea for testing this: introduce small random changes in a file, and benchmark the reparse.