Add backward compatibility for get_mla_metadata after parameter updates.

[Quan-w]

After upgrading to the latest version, the function get_mla_metadata has parameter changes:
The original parameter num_heads_per_head_k was renamed to num_q_tokens_per_head_k.

To maintain backward compatibility while supporting the new API, this PR:

1. Adds a wrapper layer in get_mla_metadata to handle both old and new parameter styles.
2. Maps the deprecated parameter num_heads_per_head_k to the new num_q_tokens_per_head_k internally.
3. Adds parameter validation (e.g., checking for conflicting parameters, missing required args, and extra unused args) to throw clear errors for invalid inputs.