fix(megatron): finalize async saves on all ranks before hooks

[HJSang]

Summary

• finalize Megatron asynchronous checkpoint saves on every training rank before invoking a post-save hook
• keep post-save hook execution restricted to rank 0
• add a regression test covering rank 0 and nonzero-rank ordering

Why

Megatron distributed-checkpoint finalization performs distributed coordination. The post-save hook path previously called blocking finalization only on rank 0, while the other ranks returned from save_model. Multi-rank jobs could therefore hang indefinitely before the hook ran.

This preserves the existing contract introduced in #1586: asynchronous saves are fully written before the hook observes them, and the hook itself runs once on rank 0. The only change is that checkpoint finalization now executes collectively.

Validation

• pytest -q --confcutdir=tests/fast/backends/megatron_utils tests/fast/backends/megatron_utils/test_actor_checkpoint.py (2 passed in radixark/miles:dev)
• pre-commit run --files miles/backends/megatron_utils/actor.py tests/fast/backends/megatron_utils/test_actor_checkpoint.py
• multi-rank reproduction: before the fix, rank 0 remained in Megatron async-call finalization after peer ranks returned; collective finalization completed normally with the fix

GPU CI

A maintainer must add run-ci-ckpt and run-ci-megatron before this draft is marked ready for review.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[HJSang]

Can someone take a look at this simple PR? it fixed async bug. Thanks!