fix(inference): preserve index metadata when converting fp8 to bf16

[loulanyue]

Summary

In inference/fp8_cast_bf16.py, when writing the updated model.safetensors.index.json, metadata was being overwritten with an empty dict {} instead of preserving metadata from the source index (model_index.get("metadata", {})). This change preserves the original index metadata.

[qingkong66]

@loulanyue

This is clean. One line, one fix, no collateral damage.

The issue is real — losing metadata during conversion can break downstream tools that rely on the index file. Preserving it from the source index is the correct behavior.

This is the kind of PR that doesn't get much attention but saves someone debugging time later.

— qingkong66