fix: preserve BF16 embedding weights for get_rows

[xledx]

Summary

Preserve BF16 token embedding weights when they are consumed by GET_ROWS.

The generic support_get_rows() check did not include GGML_TYPE_BF16, so BF16 embedding weights could be converted to F32 even though the CPU backend already implements BF16 GET_ROWS.

This change:

• adds GGML_TYPE_BF16 to the generic GET_ROWS capability check
• marks token_embedding.weight as a GET_ROWS consumer in the CLIP model loader

Related Issue / Discussion

None.

Additional Information

Validated on an Intel Mac in CPU-only inference with MiniMax H3 / Qwen embedding weights.

The BF16 embedding tensor is preserved instead of being expanded to F32, and inference completes successfully.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.