Optimize BF16 quantized matmul with FP16 tiles

[justinchuby]

Summary

• add an opt-in affine INT4/group-32 QMM path with BF16 device I/O, FP16 threadgroup/MMA tiles, and FP32 accumulation
• vectorize BF16-to-FP16 activation loads and use 64x64 output tiling
• keep AOT and Metal JIT tile instantiations consistent and test both wide/narrow variants

Enable with MLX_BF16_QMM_FP16=1; MLX_BF16_QMM_FP16_TILE=32 selects the narrow comparison path.

Performance

M1 Max, M=512, K=6656, N=19968, affine INT4 group size 32:

Path	Time
Stock BF16	30.6 ms
Stock FP16	17.2 ms
BF16 I/O + FP16 tiles	17.32 ms

The mixed path is ~1.77x faster than stock BF16 and effectively matches pure FP16 while preserving BF16 I/O.

Validation

• python/tests/test_quantized.py: 33 tests / 2896 subtests pass
• AOT and MLX_METAL_JIT=ON builds
• wide/narrow tiles, small M/N, non-64 K, and batched inputs
• JIT regression test reproduces the pre-fix tile mismatch and passes with the fix

[zcbenz]

Thanks for the PR, while it is a nice improvement we prefer users to convert to fp16 themselves rather than us implicitly do the conversion.

[justinchuby]

Thanks - explicit cast verified to have matching performance.