[Pytorch] Optimize CPU overhead for NCCL EP eager mode

[YangFei1990]

Description

Reduce NCCL EP CPU overhead by using pinned memory CPU tensor.

Fixes # (issue)

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

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

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

[YangFei1990]

/te-ci L1 pytorch

[greptile-apps]

Greptile Summary

This PR reduces eager-mode expert-parallel receive-count readback overhead by having the prepare operation write directly into pinned host memory.

• Allocates total_recv_tokens in pinned CPU memory for eager mode while retaining device storage for fixed-capacity mode.
• Explicitly synchronizes the active CUDA stream before reading the host-resident count.
• Updates API documentation and comments to describe the mode-dependent tensor placement.

Confidence Score: 5/5

The PR appears safe to merge with no actionable changed-code failures identified.

The eager path preserves the required producer-before-host-read ordering while avoiding a pageable device-to-host transfer, and the non-eager graph path retains its device-resident counter.

Important Files Changed

Filename	Overview
transformer_engine/pytorch/ep.py	Changes eager EP receive-count storage from device memory to pinned host memory and synchronizes its producer stream before sizing dispatch outputs; no actionable changed-code defect was established.

_{Reviews (1): Last reviewed commit: "Merge branch 'main' into ncclep_cpu_over..." | Re-trigger Greptile}