Add standalone fork-instrument crash reproducer

[mho22]

Summary

• add benchmarks/fork/chain.c, a recursive walker shaped after Zend's zend_compile_short_circuiting so wasm-fork-instrument's per-frame cost is measurable on a binary apart from PHP itself
• add benchmarks/fork/build.sh to produce baseline and fork-instrumented variants from one chain.c source
• add benchmarks/fork/run.mjs to binary-search the maximum recursion depth that survives V8's call-stack budget on each variant
• exit non-zero from run.mjs when forkinstr's max depth is not strictly less than baseline's — that invariant is what the benchmark exists to verify

Why

PR #669 asks for a standalone test case apart from PHP that reproduces the V8 RangeError: Maximum call stack size exceeded on deep recursion through fork-instrumented frames. Iterating on candidate fork-instrument changes via the full PHP build+run cycle takes ~45 minutes; this benchmark drops that to ~10 seconds.

The PHP crash composes two causes: unbounded recursion in zend_compile_short_circuiting (fixed PHP-side by #669) and ~8 extra wasm locals per instrumented frame from wasm-fork-instrument's switch-dispatch state machine. `chain.c` bounds recursion by its `depth` argument so it reproduces only the kandelo-side cause in isolation; any candidate reduction of the per-frame cost can be validated by watching the forkinstr depth move.

Notes

• ABI impact: none. Adds a benchmark harness only; no kernel, host, or program changes.

• Empirical numbers on macOS arm64 / Node v24.13.0, deterministic across 4 runs:

  variant	`walk` declared locals	max depth before `RangeError`	% of baseline
  baseline	4	9 972	100%
  forkinstr	12	6 648	66.7%

  Differential = 3 324 frames lost (33% drop) attributable to wasm-fork-instrument's switch-dispatch state machine adding ~8 locals per frame on the recursive function. The ratio (~1.50×) is the load-bearing finding; absolute numbers are platform/tier-dependent.

Testing

• bash scripts/dev-shell.sh bash benchmarks/fork/build.sh produces identical wasm sizes (3882 / 4587 bytes) and walk-locals counts (4 / 12) across runs
• node --experimental-wasm-exnref benchmarks/fork/run.mjs exits 0 with baseline 9 972 / forkinstr 6 648
• with the wasms manually swapped so forkinstr > baseline, the same command exits 1 with INVARIANT FAILED: ... — confirms the self-test fires on regression
• bash scripts/check-abi-version.sh exits 0 (benchmark touches no kernel/host/programs/ABI)