[None][fix] Release GIL around cudaLaunchHostFunc to fix guided-decoding host-func deadlock

[sopwg612]

Summary

launchHostFunc held the Python GIL across cudaLaunchHostFunc. The CUDA driver serializes host-function dispatch on a stream, so on the eager (non-capturing) path a launch issued while a callback is in flight busy-waits in the driver. That in-flight callback runs cudaHostFuncTrampoline, which reacquires the GIL to execute Python (e.g. xgrammar in guided decoding). With the launching thread holding the GIL, the callback can never acquire it -> circular deadlock, zero forward progress, no crash.

Triggered by Eagle3 speculative decoding + xgrammar guided decoding. Captured live via py-spy --native on both TP ranks of a wedged engine:

• Thread-3 (holds GIL): sched_yield -> libcuda -> launch_hostfunc -> GuidedDecoder.execute -> eagle3.forward
• host-func worker: ~gil_scoped_release (reacquiring GIL) -> xgrammar fill_next_token_bitmask -> GuidedDecoder.build

Root cause

Regression from NVIDIA#13251, which correctly removed the call_guard (so the nanobind arg construct/destruct happens under the held GIL) but also removed the inner GIL release around the driver enqueue that NVIDIA#10028 had added.

Fix

Re-release the GIL only around cudaLaunchHostFunc, and only on the eager path (freeUserData == true). During graph capture no callback runs, so the GIL is kept to hold capture on a single thread. The release scope ends before the nanobind refcount work, preserving NVIDIA#13251's object-lifetime fix (nvbugs-5643631). Reached only via CapturableGuidedDecoder (one-model spec + guided); non-guided / non-spec paths are untouched.

Validation

• Root-caused on a live ~25h-wedged repro (gpt-oss-120b-Turbo, tp2/ep2) via py-spy --native; both ranks show the identical cycle above.
• Adds a regression guard test_hostfunc_gil_release_no_deadlock. NOTE: not executed in this PR (needs a GPU + a C++ rebuild); the production repro is the primary validation.
• Interim mitigation for deployed rc16 without this build: disable Eagle3 spec-decode (removes the host-func guided path).

Scope

Fork hotfix targeting deep-main-v1.3.0rc16. The same patch is suitable for NVIDIA upstream (main) as a follow-up.