perf(ios): add engine native dispatch fast paths

[DjDeveloperr]

Summary

• Adds engine-direct native dispatch paths for V8, JavaScriptCore, QuickJS, and Hermes so native API calls can stay on engine-native callbacks and avoid the generic Node-API dispatch path where possible.
• Keeps gsd-off on the same engine-direct callback/marshalling path for every engine; only generated typed invoker lookup is disabled, so the comparison isolates GSD.
• Tightens generated/native object dispatch with fast block invoke, backend-specific generated converters, direct fallback for non-generated signatures, lighter class/protocol wrapper normalization, a JSC unwrap cache, and a Hermes receiver cache.
• Keeps the generic Node-API path available for non-direct/non-target builds while V8/JSC/QuickJS/Hermes builds use their engine-specific direct layers.

Benchmarks

Simulator: iPhone 17 Pro Max. Harness: npm run benchmark:objc-dispatch, 250k base iterations. V8 generic NAPI backend is intentionally not included.

Runtime / mode	Total ms	Delta vs off	Ratio vs off	Notes
NAPI iOS V8 GSD on	339.38	-101.34	1.30x faster	fastest run
NAPI iOS V8 GSD off	440.71	baseline	1.00x	engine-direct path, GSD disabled
NAPI iOS QuickJS GSD on	459.35	-52.42	1.11x faster	faster than legacy V8 AOT
NAPI iOS QuickJS GSD off	511.77	baseline	1.00x	engine-direct path, GSD disabled
NAPI iOS JSC GSD on	502.32	-52.86	1.11x faster	faster than legacy V8 AOT
NAPI iOS JSC GSD off	555.18	baseline	1.00x	engine-direct path, GSD disabled
Legacy iOS V8 AOT on	812.04	-45.67	1.06x faster	exact same harness
Legacy iOS V8 AOT off	857.71	baseline	1.00x	exact same harness
NAPI iOS Hermes GSD on	1368.02	-562.49	1.41x faster	GSD helps, engine boundary still dominates
NAPI iOS Hermes GSD off	1930.51	baseline	1.00x	engine-direct path, GSD disabled

Result files used for the final combined report:

• build/benchmarks/objc-dispatch/results-2026-05-22T03-50-08-551Z.json Hermes
• build/benchmarks/objc-dispatch/results-2026-05-22T03-51-20-831Z.json JSC
• build/benchmarks/objc-dispatch/results-2026-05-22T03-52-29-898Z.json QuickJS
• build/benchmarks/objc-dispatch/results-2026-05-22T03-54-25-288Z.json V8
• build/benchmarks/objc-dispatch/results-2026-05-22T03-54-53-235Z.json legacy iOS V8 AOT on/off
• build/benchmarks/objc-dispatch/results-combined-final-2026-05-22.json combined harness report

Notes

• GSD on/off is now visibly separated across every direct backend: V8 +30%, QuickJS/JSC +11%, Hermes +41%.
• V8 direct + GSD is now well ahead of legacy iOS V8 AOT on/off in this harness.
• QuickJS and JSC direct + GSD are also ahead of legacy iOS V8 AOT on/off.
• Hermes improves substantially with GSD but still has a much higher floor than V8/QuickJS/JSC; the remaining cost appears to be Hermes host-function/wrapper entry overhead rather than libffi/generated-signature dispatch.

Verification

• git diff --check -- <touched runtime files> exits 0.
• Clean detached worktree with only this final patch applied: npm run test:ios passes with tests=705, failures=0, errors=0, skipped=14 on iPhone 17 Pro Max simulator.