Add Metal paged attention minimal support (flagged) and SDPA binding fix (OPTIONAL fast path for Continuous Batching PR)

[Sohailm25]

Proposed changes

• Add minimal Metal paged attention support with a Python fallback and keep SDPA as the default. This enables downstream consumers (e.g., mlx-lm) to opt into paged attention without changing behavior for
  existing users.
• Align SDPA bindings with current mlx main (add is_training / output_logsumexp, restore VJP / is_equivalent symbols) to fix build/link drift.
• Tests cover paged KV allocation/copy, fallback parity, and SDPA fast-path (python/tests/test_paged_kv.py, python/tests/test_fast.py).

Checklist

• I have read the CONTRIBUTING (https://github.com/ml-explore/mlx/blob/main/CONTRIBUTING.md) document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[Sohailm25]

This adds future pathway for the Continuous Batching change in mlx-lm