[Dev] Fix full CUDA graph capture reverted by pull main

[Victarry]

Summary

• Restore the source-rank cu_seqlens length broadcast allocation to torch.empty(1, dtype=torch.int64, device=dev).fill_(n).
• Avoid torch.tensor(n, dtype=torch.int64, device=dev), which can perform a CPU-to-CUDA copy and fails during full-iteration CUDA graph capture when the CPU source is not pinned.
• This restores the earlier dev behavior from f9f9fe415 (Add minor tweaks to support full-iter CG capture). The nightly pull-main chore: nightly sync main into dev (10_05_2026) #4716 post-merge commit cd2001de7 (chore: post-merge fixes for nightly sync main into dev (28_04_2026)) took main's megatron/training/utils.py and reverted this line back to torch.tensor(...).

Test plan

• python -m py_compile megatron/training/utils.py
• CHECK_ONLY=true BASE_REF=dev uv run --project /Users/denliu/Projects/repos/agentic-mcore-dev --extra mcore-lint -- bash tools/autoformat.sh
• Copyright check: no megatron/core or tests Python files changed.
• Ran paged-stash dev config on Computelab B300 (umb-b300-003) with local/b300-cudnnfe1221:latest; reached iteration 100, printed validation loss, and no longer hit the CUDA graph capture error from torch.tensor(n, device=dev).

[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.

[Victarry]

/ok to test 05fc772