[PyTorch] Fix GIL/refcount abort in Comm+GEMM overlap and NCCL-EP init bindings

[denera]

Description

The factory py::init bindings for CommOverlap/CommOverlapP2P and the ep_initialize binding were registered with an outer py::call_guard<py::gil_scoped_release>(), which releases the GIL for the entire call -- including regions where pybind11 performs Python reference-count operations that require the GIL to be held. This triggers a pybind11 GIL/ref-count hazard assert when TE is compiled with NVTE_BUILD_DEBUG=1.

PR fixes the bug by dropping the outer call guard and releasing the GIL narrowly, only around the blocking C++ native constructor/initializer. The GIL is reacquired before pybind11 casts the arguments, registers the holder and destroys the by-value object parameters.

Fixes # 3199

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

This PR narrows where the PyTorch bindings release the GIL.

• ep_initialize now keeps the GIL during pybind argument handling and releases it only around native EP initialization.
• CommOverlap and CommOverlapP2P constructors now release the GIL inside the init lambda around native construction.
• The outer pybind call guards were removed from the affected init bindings.

Confidence Score: 5/5

This looks safe to merge.

• No blocking issues found in the changed code.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/csrc/extensions/ep.cpp	The EP init binding now releases the GIL only around the native initializer while keeping pybind object handling under the GIL.
transformer_engine/pytorch/csrc/extensions/pybind.cpp	The CommOverlap bindings now release the GIL inside the constructor lambdas after pybind has converted the Python arguments.

_{Reviews (1): Last reviewed commit: "Fix GIL/refcount abort in Comm+GEMM over..." | Re-trigger Greptile}

[timmoon10]

LGTM