Add weight-coverage walker to converter test suite

[jlamypoirier]

Summary

• Extends tests/models/test_converters.py with a per-fixture weight-coverage check: for each ModelTestingConfig with a checkpoint_format, materialise the Fast-LLM base model and assert every named_parameters() entry is consumed by some leaf WeightConverter emitted by base_model_converter_class.get_converters(config).
• Tied parameters (e.g. MTP head output weights sharing a single LM-head weight) count as covered when any group member has a converter, matching the export-time behaviour where one shared weight is serialised once.
• gemma4 is xfailed (strict) against pre-existing coverage gaps in its converter declarations; the xfail reason describes the three specific gaps.

Why

The strict HF-side coverage check landed by #523 catches an HF dict carrying a key with no Fast-LLM consumer. There was no equivalent symmetric check on the Fast-LLM side — a model parameter without a converter is silently skipped by ExternalStateDictCheckpointHandler._convert_state_dict, so its trained value is lost on HF export. This test closes that gap statically.

Implementation notes

• Uses bm.base_model_class(bm, DistributedConfig()) — CPU-only, ParameterMeta weights, no distributed setup or NCCL. Iterates named_parameters() for the canonical Fast-LLM name set.
• Tied-group resolution uses object identity (id(parameter)) since ParameterMeta.tensor_name isn't assigned until stage setup, which we deliberately skip.
• Walker is parametrised over every fixture with a checkpoint format (not just one per format) so format-specific variations get coverage. Cheap — adds ~36 tests, all under 0.7s combined locally.

Test plan

• Local: pytest -v tests/models/test_converters.py → 36 pass, 1 xfail (gemma4)
• Cluster: full tests/models/test_converters.py + test_checkpoint.py + test_hf_roundtrip.py sweep — converters tests pass + xfail as designed
• Pre-existing failures (test_conversion[*] torchscript coverage and test_hf_roundtrip Gemma4 import) verified unrelated to this PR (diff is +71 lines in test_converters.py only)

🤖 Generated with Claude Code