Avoid static stream registry teardown

[voltjia]

Summary

This PR keeps the process-wide stream registry alive until process termination so extension static teardown cannot call InfiniRT after the accelerator runtime has started shutting down.

It removes the native stream and event destruction loop that was observed hanging inside CUDA runtime teardown and relies on the operating system to reclaim process-scoped handles.

It adds a subprocess regression that creates multiple streams, submits Add work, synchronizes the streams, and requires Python to exit within 30 seconds.

Validation

• Ruff and clang-format 19.1.4 checks passed.
• The installed wheel built and imported with PyTorch 2.13.0+cu130 on the NVIDIA runtime backend.
• The subprocess exit regression passed in five independent NVIDIA containers.
• The complete installed-wheel CPU fallback suite passed with 142 tests passed, 32 skipped, and 3 expected failures.
• The broader NVIDIA stream suite completed without teardown hangs. Five existing Add value assertions still fail with the cached SM75 InfiniOps build on A100 and are tracked separately in Stabilize NVIDIA Add validation #33.