Noir::FastSearch

A memchr-driven fast path for String#index(String), String#byte_index(String) and therefore String#includes?(String).

The stdlib implements substring search as a Rabin-Karp rolling hash: one multiply-add per byte, no vectorisation, roughly 0.8 GB/s. Every detector and analyzer gates its regexes with content.includes?("..."), so on a scan that reads tens of megabytes of source the search itself was a measurable share of the wall clock (profiled on superset: ~9% of the Flask analyzer, ~50% of the Express analyzer).

The replacement scans for the rarest byte of the needle with memchr (vectorised in every libc) and confirms each candidate with memcmp. It tests every byte offset, so a miss here is a miss in the stdlib too. A hit is returned directly only when the stdlib would have found the same one: for byte_index always (both are byte-level), for index only when every byte before the hit is ASCII, because the stdlib steps by decoded character and can skip over a match that begins inside a multi-byte or malformed sequence. Every other case falls through to previous_def, so the result is identical by construction rather than by argument. The fall-through is bounded by the old cost: it only runs after a hit, and the stdlib scan stops at that same hit.