[NFC] Improve SimplifyLocals compile-time performance

[Changqing-JING]

The checkInvalidations function in SimplifyLocals had O(N × S) complexity, iterating all sinkables for every visited expression. This PR adds reverse indices (localReadBySinkable_, localWrittenBySinkable_, heavySinkables_) so that checkInvalidations only examines sinkables whose effects actually conflict with the current expression. The Sinkables map and localsRead/localsWritten sets are also changed from ordered to unordered containers for faster lookups. A new hasNonLocalOrderingEffects() helper is added to EffectAnalyzer (composing existing helpers like writesGlobalState() and readsMutableGlobalState()) so the classification stays in sync automatically when new effect types are added.

Benchmark data
Use the wasm file in
#7319 (comment)

Note: don't run benchmark with -Oz, use --simplify-locals can save time.
Before this PR

time ./build/bin/wasm-opt --simplify-locals --enable-bulk-memory --enable-multivalue --enable-reference-types --enable-gc --enable-tail-call --enable-exception-handling  -o /dev/null ./test3.wasm

real    0m42.423s
user    9m45.063s
sys     0m44.099s

In this PR

time ./build/bin/wasm-opt --simplify-locals --enable-
bulk-memory --enable-multivalue --enable-reference-types --enable-gc --enable-tail-call --enable-e
xception-handling  -o /dev/null ./test3.wasm

real    0m29.725s
user    6m25.124s
sys     0m30.590s

[kripken]

IIUC this tries to be faster by guessing that most interactions are local?

If that's the case, then I don't think that is true in general. We would need to measure on more codebases to check that. If it isn't generally true, then it could make us slower sometimes. It also adds chances for bugs to crop up here, as this code must remain in sync with effects.h.