fix: gpt-oss ckpt saving

[akoumpa]

What does this PR do ?

1. Root cause fix in _maybe_build_consolidated_index: filters the final fqn_to_file_index_mapping to only include keys actually present in the state_dict being saved. This removes phantom _blocks/_scales keys from the base checkpoint index when saving dequantized bf16 weights.
2. Defensive fix in _parse_input_metadata: after populating output file data from input shards, removes any FQNs that still have an empty dtype_str (meaning they weren't found in any input shard). Logs a warning to make the mismatch visible.
3. Four new tests:

• TestBuildConsolidatedIndexQuantizedBase::test_stale_quantized_keys_are_filtered - verifies _blocks/_scales keys don't leak into the mapping
• TestBuildConsolidatedIndexQuantizedBase::test_new_keys_assigned_to_last_shard - verifies new keys get the default shard index
• TestConsolidationWithPhantomKeys::test_consolidated_safetensors_loadable_with_phantom_keys - end-to-end: creates DCP shard files, runs consolidation with phantom keys, verifies the output is loadable with safe_open and contains correct tensors
• TestConsolidationWithPhantomKeys::test_consolidated_safetensors_no_phantom_keys - regression test: consolidation without phantom keys works as before

Changelog

• Add specific line by line info of high level changes in this PR.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?

If you haven't finished some of the above items you can still open "Draft" PR.

Additional Information

• Related to # (issue)

[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.

[akoumpa]

/ok to test 4d4fc31

[akoumpa]

/ok to test 8c6826b

[akoumpa]

/ok to test b5f2606

[akoumpa]

/ok to test f281b65