Fix Python finfo.min port in quantized attention masking

[ronaldmannak]

Proposed changes

Fixes a likely mistranslation from Python mlx-lm in quantizedScaledDotProductAttention.

The Python reference masks boolean/causal logits with mx.finfo(scores.dtype).min, a large negative finite value that suppresses masked positions before softmax. The Swift port used Float.leastNormalMagnitude, which is instead a tiny positive value, so masked positions could remain competitive or dominate when valid scores are negative.

This updates the Swift quantized attention mask fill value to match the intended Python semantics and adds regression coverage for masked quantized attention.

Checklist

Put an x in the boxes that apply.

• 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)

[ronaldmannak]

That turned out to be a little bit more complex than I thought, but should be good to go now. Also in this PR: a related CQA fix in the same KVCache method

[ronaldmannak]

One open question: I added extensions to MLXArray and DType to this repo. You could argue those belong in MLX-Swift. Happy to open a separate PR on MLX-Swift for that

[davidkoski]

One open question: I added extensions to MLXArray and DType to this repo. You could argue those belong in MLX-Swift. Happy to open a separate PR on MLX-Swift for that

Yeah, that would probably be the best approach. There is a very thin finfo on there already

[ronaldmannak]

@davidkoski Done: ml-explore/mlx-swift#429