Cache a protocol member's parsed requirement; Fix #363

[enumeratio]

Fix #363

requirementShape and requirementAt re-parsed the member's signature on every dispatched call. A signature naming Self never enters parseType's shared cache, and its resolver-less first attempt throws before the resolver-aware parse runs. The parsed requirement is now kept per protocol record (a WeakMap), keyed by member and the receiver's Self spelling, and the entry is dropped when ce._cacheGeneration() moves, so a redeclared type re-parses. A read-only engine view without a generation parses as before.

Measured on 3,000 scalar-receiver calls: 421 ms → 41 ms, against 32 ms for an equivalent plain function.

[enumeratio]

Pushed a review pass:

• The cache is keyed on the engine's any version rather than _cacheGeneration(). The low bit of _cacheGeneration() flips while assumptions are hidden, which would empty the entry on every toggle, and what a requirement parses to doesn't depend on assumptions.
• It's bounded per protocol (REQUIREMENT_CACHE_LIMIT). A generic conformance meets a new Self for every list length it's called with, so an unbounded entry grows with the calls rather than the declarations.
• Added regression tests (item-363-protocol-requirement-cache.test.ts):
  • repeated calls over interleaved receiver types answer what a fresh parse answers, rejections included;
  • a conformance added after a call serves the next call;
  • 300 distinct receiver types keep answering past the bound.

Member calls on the benchmark from #363 stay about 11× faster than on main. The PR also merges cleanly with #366 (CHANGELOG aside), and the combined result is just as fast.