fix: review findings #1, #2 from PR #1126

[adobrzyn]

Fixes from code review of #1126:

• Add pre-commit GHA #1: Add unit tests for _gather_mm_embeddings covering single-request, two-request contiguous, and 2D padded batching paths
• Refactor vllm_hpu -> vllm_gaudi #2: Use separate local variable (effective_total_tokens) instead of overriding the total_num_scheduled_tokens parameter in _get_model_mm_inputs

---

[- Reviewed by Awesome ChlOpus]