feat(bench): add WASM benchmark harness and native allocation profiler

[k4sperski]

Issues

Follows up on #3025.

Description

Adds benchmarking infrastructure to measure spell-check performance in WASM and track allocation overhead; two dimensions native Criterion benchmarks don't cover. fuzzy_match is cheap on modern machines, but on lower-end devices WASM is slower and allocation costlier; these tools make that overhead visible.

Validated against #3025 (is_already_lower):

Metric	Lowercase	Mixed	Capitalised
WASM time reduction	49.2%	43.4%	~0%
Allocation churn reduction	44.7%	38.9%	~0%

WASM time reductions track #3025's native Criterion numbers, so the optimisation carries through the WASM pipeline. Capitalised is flat because is_already_lower doesn't trigger on uppercase starts.

What's added

• Native allocation profiler (harper-core/examples/alloc_profile.rs): counts allocs per word in fuzzy_match via a #[global_allocator]. No extra deps. Run with just alloc-profile.
• WASM bench wrapper (harper-wasm/src/bench.rs): feature-gated PreparedWords struct exposing fuzzy_match to JS. Constructor parses the word list and caches an Arc<FstDictionary>; the method just runs the spell-check loop. cfg(feature = "bench") keeps it out of production builds.
• JS bench harness (harper-wasm/benches/wasm_bench.js): Node script timing fuzzy_match under WASM to verify native perf wins carry through. Run with just bench-wasm.

Measurement hygiene

• PreparedWords parses inputs and caches the dictionary once, outside bench(), so the timed region only contains fuzzy_match work, matching the native criterion pattern.
• The alloc profiler warms the AUTOMATON_BUILDERS thread_local before measuring, so the first case doesn't absorb its one-time lazy-init cost.
• `MAX_EDIT_DISTANCE = 3 and MAX_RESULTS = 200 are duplicated across the alloc profiler, the criterion bench, and the JS harness; a comment flags the invariant.
• just bench-wasm builds into pkg-bench/, so the shipping pkg/ stays untouched

WASM Time: baseline vs optimised

xychart-beta
    title "WASM fuzzy_match: ms/iter min (lower is better)"
    x-axis ["mixed", "lowercase", "capitalized"]
    y-axis "ms/iter" 0 --> 100
    bar [96, 84, 80]
    bar [54, 43, 80]

Loading

Methodology

• baseline = early-exit guard removed, optimised = guard restored. Numbers are the min of 50 iterations (after 5 warmup iterations) measured on the default just bench-wasm build (wasm-opt on). Same word lists and parameters as the native Criterion bench

Trade-offs

• JS harness isn't in CI, so value depends on contributors running just bench-wasm before merging perf-sensitive changes. CI integration would be the natural next step if this becomes a recurring gap.
• Running bench-wasm with the default wasm-opt build adds ~30s per rebuild. DISABLE_WASM_OPT=1 just bench-wasm skips it for faster dev iteration; the numbers that ship in this PR were measured with wasm-opt on.

Follow-ups

• If the harness graduates to CI, pass MAX_EDIT_DISTANCE / MAX_RESULTS as args so just bench-wasm becomes the single source of truth for the duplicated constants.
• Extend PreparedWords to cover suggest_correct_spelling and merged-dict lookup when optimisation work warrants it.

Demo

N/a

How Has This Been Tested?

• cargo build -p harper-wasm: production build unaffected (no bench feature).
• cargo clippy -p harper-wasm --features bench -- -Dwarnings: clean.
• just alloc-profile: first case no longer inflated by thread-local init.
• just bench-wasm: runs end-to-end; PreparedWords.free() called so no wasm-bindgen leaks.

Checklist

• I have performed a self-review of my own code
• I have added tests to cover my changes
• I have considered splitting this into smaller pull requests.

[k4sperski]

Moving to draft, needs a bit more refinement!