Drop the unreachable re-entry guard from _repo_lock

[anandhu-eng]

Stack: #265 ← #323 ← #325 ← #326. Merge in that order: #323 into #265's branch first, then each next PR. This PR targets #325's branch. All 125 tests pass at the top of the stack.

Why

#265 has two thread-local sets that both mean "this thread already has this repo":

	_held_repo_locks (in _repo_lock)	_pulling_repos (in _repo_pull_lock)
On re-entry	Continues without locking, and redoes the work	Skips the work and returns {'return': 0}
Reached today	No	Yes, on dependency cycles

Both came from #303. The first fixed a self-deadlock: register_repo pulls each dependency while the parent repo's lock is held, and a second FileLock on the same file blocks even within one thread. But that fix turned a dependency cycle into infinite recursion, so the second set was added to skip the nested pull.

The first set is now unreachable:

• _repo_pull_lock checks _pulling_repos before it calls _repo_lock, and a path is only in _held_repo_locks while it is also in _pulling_repos.
• _repo_lock's only other caller, RepoAction.rm(), never runs inside a pull.
• Removing the check fails only the one test that called _repo_lock directly. No pull or rm test notices.

Changes

• _repo_lock is now a plain lock. _held_repo_locks and _repo_locks_held_by_this_thread are removed.
• The note about two threads pulling repos that depend on each other (a real lock-order inversion that still ends only on timeout) moves onto _pulling_repos. Comments that described same-thread re-entry as a no-op are corrected.
• Tests:
  • test_same_thread_reentry_does_not_self_deadlock is removed. It only exercised the deleted branch.
  • test_dependency_cycle_does_not_recurse now also asserts that the nested entry doesn't block. With the remaining guard removed it fails with a lock Timeout, which I checked.

No behaviour change. Locally: 120 unit tests pass, plus .github/scripts/test_repo_pull_force.py.

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[github-actions]

🤖 AI PR Review Summary

This PR removes the previously used thread-local set for tracking held repo locks and the reentrancy logic in _repo_lock, simplifying the locking mechanism. Instead, it relies on _repo_pull_lock to guard against same-thread recursive pulls. The comments have been updated to clarify the locking behavior and limitations. The test for same-thread reentry deadlock was removed, reflecting the change in locking strategy. Overall, the changes reduce complexity but remove the within-thread reentrancy in _repo_lock, which may affect code relying on that behavior. The tests verify that locks are released properly and that dependency cycles are guarded by _repo_pull_lock. No new inline actionable issues found in the diff.