Share compiled statement lists across generator and async invocations

[lahma]

Follow-up to #2567. That fix gave each generator instance its own JintStatementList to stop live instances from corrupting each other's resume position — at the cost of rebuilding the whole compiled statement handler tree on every generator instantiation (a price async statement bodies had also been paying per invocation since their introduction).

Change

JintStatementList is now immutable after construction: the mutable resume position (_index) moves into the suspendable instance's SuspendDataDictionary, keyed by the list — completing the pattern already used for every other piece of suspension state (loop iterators, destructuring, block environments). With positions per instance, the cached list is safe to share, so:

• sync and async generator bodies reuse the cached statement list again (_bodyStatementList ??=),
• async function statement bodies stop allocating a fresh list (and handler tree) per call.

Most suspendables only ever track one list (the function body), so a one-entry inline slot fronts the position dictionary; nested-list suspensions fall back to the dictionary. A key never moves between slot and dictionary, so lookups cannot observe stale entries. The sync execution path only gains one frame-suspendable read per list execution — it was already read per statement for the suspension checks.

Benchmarks

BenchmarkDotNet default job, .NET 10, A/B against main (059f174) from separate worktrees, idle machine. GeneratorBenchmark is added in this PR (small bodies, a large body with nested blocks/loops/switch, and interleaved instances).

Benchmark	main	PR	Δ time	Δ alloc
SmallGenerators_1000	1,310.0 us / 2,640 KB	1,247.4 us / 2,417 KB	−4.8%	−8.5%
LargeBodyGenerators_500	3,725.6 us / 7,813 KB	2,711.7 us / 3,010 KB	−27.2%	−61.5%
InterleavedGenerators_500	1,166.3 us / 1,752 KB	1,150.9 us / 1,382 KB	−1.3%	−21.1%
PlainAsyncFunctionExit_1000	556.3 us / 2,033 KB	461.9 us / 1,802 KB	−17.0%	−11.4%
AsyncFunctionWithSyncUsing_1000	961.4 us / 2,818 KB	768.1 us / 2,263 KB	−20.1%	−19.7%
PlainAsyncGeneratorExit_500	1,453.1 us / 4,955 KB	1,349.2 us / 4,843 KB	−7.1%	−2.3%

Sync neutrality probes (StopwatchBenchmark ×4, RecursionBenchmark Fib/Tak/DeepSum): allocations byte-identical to main on all seven; time deltas ranged −4.9%..+6.4% with inconsistent signs across closely related methods (e.g. Stopwatch Execute +5.3% while Execute_ParsedScript −2.8%), i.e. run-to-run noise rather than a systematic shift.

Verification

• Jint.Tests: 3160 (net10.0) / 3098 (net472) passed, 0 failed — including the Fix async function parameter binding after await resumption #2567 interleaved-generator regression tests (top-level, nested-block, for..of termination, async generators)
• Jint.Tests.PublicInterface: 79 + 79 passed
• Test262: 99260 passed, 0 failed (baseline parity)

🤖 Generated with Claude Code