[codex] Add experimental FSDP meta-parameter support

[wujingyue]

This is intentionally separated from the minimal FSDP cleanup in the current stacked branch so the meta-parameter support can be kept, revised, or dropped independently.

What changed:

• Restores experimental fully_shard handling for direct meta parameters by materializing with to_empty(...) and invoking reset_parameters() / _reset_parameters() before ParameterGroup construction.
• Restores the distributed test that verifies meta parameters are replaced by initialized sharded DTensors.
• Keeps the CPU-initialized real-parameter coverage added in the base branch.
• Does not touch non-experimental FSDP files.

Validation:

• CUDA_VISIBLE_DEVICES= uv run --no-sync python -m torch.distributed.run --nproc-per-node 8 -m pytest -q tests/unit_tests/distributed/megatron_fsdp/test_experimental_fully_shard.py (8 passed, 1 skipped)

Notes:

• The plain requested uv run ... command could not run in this session because uv attempted to modify /opt/venv and hit a permission error.
• Running with CUDA visible could not complete on this host because it has 2 GPUs while the test command launches 8 local ranks.

[wujingyue]

Moved to upstream PR NVIDIA#5369.