From #630:

We should cache the extensions/capabilities into a hashset or something instead of doing a linear search every time (and string cmp on top of that). There's probably a lot of other inefficiencies like this littered around the codebase, too, but yeah.

(Also we should really look into removing/bypassing rspirv's type deduplication code, it's wasting quite a bit of work double-deduplicating stuff after our deduplication - and their deduplication is a linear scan through all types_global_values, unlike our hash lookup)