[Dev] Add combined 1F1B schedule-aware tensor release

[lhb8125]

Summary

• Track producer streams on concrete tensor objects within each combined 1F1B model-chunk plan.
• Consume each binding at the next real node: free_input=False only drops ownership metadata, while external or unregistered tensors retain the conservative record_stream fallback.
• Export tensors that cross plan boundaries and drain deferred releases during phase finalization, with an audit for dangling owner bindings.
• Reject all storage aliases in managed node transitions so object-level ownership and EMPTY_STORAGE remain unambiguous.
• Keep the detailed lifetime-management documentation next to the implementation in combined_1f1b_tensor_release.py, and scope implementation names to this schedule.

This supersedes #6948 with the reviewed tensor-bound ownership model.

Test plan

• Project Python py_compile on all 12 changed Python files.
• CHECK_ONLY=true BASE_REF=dev bash tools/autoformat.sh (Black, isort, Pylint, and Ruff pass).
• python tools/check_copyright.py on the complete PR changeset.
• CMH GB200 CI container: tests/unit_tests/a2a_overlap/test_combined_1f1b_tensor_release.py — 21 passed in 4.52s.
• The allocator functional test compares identical cross-stream dependencies and verifies scheduled release reuses the owner allocation and lowers peak reserved memory versus record_stream while the consumer is still in flight. It takes 0.12s; total file time increased from 4.41s to 4.52s.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Wohox]

What about setting to True by default? Maybe we need more benchmarking data to support this comment.

[lhb8125]

Good point. I am keeping this opt-in for now because the current evidence covers correctness and a native-allocator microbenchmark, but not end-to-end throughput and peak memory across the supported EP-overlap configurations. This knob is only useful for the sync-free combined 1F1B path with overlap_moe_expert_parallel_comm=True; other schedules can keep the default allocator record_stream path. Also, changing this field to True alone would make the default TransformerConfig invalid because validation requires the parent overlap flag. I clarified when users should enable the knob in 188e5f5b4. We can revisit a conditional default-on behavior, with an explicit opt-out, after representative on/off benchmarks.

[lhb8125]

This knob only helps the sync-free path, which is not widely used.

[Wohox]

Make sense, but for non sync-free path where CPU is faster like DSv4/3 case this toggle should also help?

[lhb8125]

/ok to test 930c99c

[lhb8125]

/ok to test 3a8b253

[lhb8125]

/ok to test 7dc1595

[Wohox]

No more comments, LGTM!