The problem does not affect v3.

This should simplify the logic and significantly improve
performance in edge cases as found and reported on #537
by CJ Cullen.

This change introduces an index to lookup token numbers referenced by simple_keys in O(1),
thus significantly reducing the performance impact of certain abusively constructed snippets.

When we build up the simple_keys stack, we count on the (formerly named) staleness check to
catch errors where a simple key is required but would be > 1024 chars or span lines. The previous
simplification that searches the stack from the top can go 1024 keys deep before finding a "stale"
key and stopping. I added a test that shows that this consumes ~3s per 1MB of document size.