Fix matmul precision with broadcast RHS

[Sarves1911]

Proposed changes

Fixes #3953.

The Metal matmul backend was collapsing a broadcast batch dimension into M even when the original operation had M == 1.

For a float32 batched matrix-vector product with a stride-zero broadcasted right-hand operand, this changed the operation from the GEMV path to the TF32-enabled GEMM path. As a result, multiplying by exactly 1.0 could reduce the precision of the right-hand operand.

This change prevents the batch-into-M collapse when the original M dimension is 1, allowing the operation to remain on the existing GEMV path.

A regression test is included that runs the operation with TF32 enabled in a separate process.

Testing

• Focused regression test passed:
  PYTHONPATH=python/tests python -m unittest test_blas.TestBlas.test_matmul_broadcast_rhs_k1_precision -v
• Complete BLAS test suite passed:
  PYTHONPATH=python/tests python -m unittest test_blas -v
• 29 BLAS tests passed.
• Pre-commit hooks passed for the modified files.
• The original issue reproduction now reports a maximum error of 0.0 with MLX_ENABLE_TF32=1.

Checklist

• I have read the CONTRIBUTING 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)