Validate MFSDP module phase transitions

[wujingyue]

What does this PR do?

Make MFSDP v2 module-phase handling explicit and validated after #6153:

• Add a public FsdpModule.phase property that permits only RESTING → FORWARD, FORWARD → RESTING, RESTING → BACKWARD, and BACKWARD → RESTING.
• Route forward and backward hooks through that property instead of writing _phase directly.
• Finalize the MFSDP context in pre_forward(), the first MFSDP hook to run.
• Separate activation-recomputation detection from the phase transitions. Reentrant recomputation follows RESTING → FORWARD → RESTING; non-reentrant recomputation occurs between pre_backward() and post_backward() and preserves BACKWARD through its nested forward hooks. In both cases, recomputation skips forward-order prefetch and parameter resharding.
• Update the activation-recompute test to assert the public phase state after backward and after a subsequent forward/backward pass.

Validation

• python -m torch.distributed.run --nproc-per-node 2 -m pytest -q tests/unit_tests/distributed/mfsdp_v2/test_fully_shard.py -k activation_recompute
• ruff check megatron/core/distributed/fsdp/src/megatron_fsdp/experimental tests/unit_tests/distributed/mfsdp_v2/test_fully_shard.py

[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.

[wujingyue]

/ok to test 652fbe1