Account for fast-glob vs chokidar glob differences

fast-glob and chokidar understand globs in slightly different ways, plus we currently do a bunch of extra normalization etc. to the globs we pass to fast-glob, which we don't do to chokidar.

Ideally the behaviors would be identical.

It's actually OK if chokidar sometimes handles more files than necessary, since we only use chokidar to notice a change *anywhere*, and always rely on fast-glob to figure out what *actually* changed for a given iteration (which could end up being nothing). A difference here would likely indicate a performance issue, though, if we're watching more than we need to.

It's not OK if chokidar sometimes misses files that fast-glob does, since that means we won't trigger an iteration in all cases when we should.