PERF: release the GVL while the V8 thread boots the isolate

[ursm]

Summary

In the default multithreaded mode, Context.new spawns the V8 thread and then blocks on the late_init barrier while that thread runs v8::Isolate::New (snapshot deserialization) and the safe-context setup. barrier_wait is a plain pthread_cond_wait, so the booting Ruby thread holds the GVL for the entire boot, freezing every other Ruby thread for the full duration.

The V8 thread touches no Ruby objects during boot — the snapshot is a plain C buffer already copied into c->snapshot — so this change releases the GVL around the barrier waits via rb_thread_call_without_gvl (the same idiom already used by Context#dispose). A background thread booting contexts from a snapshot (e.g. a warm-context pool) now genuinely overlaps with the main thread instead of stalling it.

Why it's safe

• The boot path (early_init → v8::Isolate::New / context / safe-context setup → late_init) makes no Ruby calls and allocates no VALUEs before late_init, so the initializing thread never needed to hold the GVL there.
• self/c stay alive across the GVL-released wait via MRI's conservative machine-stack marking — identical to the existing rb_thread_call_without_gvl(context_dispose_do, c, NULL, NULL) in Context#dispose.
• context_mark only reads c->procs / c->exception, both initialized before the V8 thread exists and untouched during boot, so a concurrent GC mark is race-free.
• Only the multithreaded branch is affected; :single_threaded mode (which runs the boot inline on the Ruby thread) is unchanged.

Measurement

Synthetic microbenchmark (background thread booting contexts from a heavy snapshot while the main thread runs a tight Ruby loop): on main, every background boot stalls the main thread for the full ~8 ms deserialization; with this change those per-boot stalls disappear.

Real-world A/B in a downstream app that maintains a warm Context pool with background async refill (real snapshot + pool + refill, median of 8 runs, identical workload): ~1.52s → ~1.40s wall, ≈ 8.5% faster.

All existing tests pass (rake test, 113 runs, 0 failures).

🤖 Generated with Claude Code