Use unroll_count(4) for the NAX attention Q@K.T loop

[wyanzhao]

Proposed changes

Change the NAX attention Q @ K.T head-dimension loop from full unrolling
to #pragma clang loop unroll_count(4).

Note

This is a scheduling-only change. The accumulation order and numerical
results are unchanged.

Motivation

With full unrolling, the compiler hoists all TD K-tile loads to the start
of the block, ahead of the running MMA chain. Unrolling by four instead lets
the compiler interleave K-tile loads with matrix accumulation.

Configuration	Full unroll	unroll_count(4)
head_dim == 128 (TD == 8)	All K-tile loads are hoisted	Loads can interleave with the MMA chain
head_dim == 64 (TD == 4)	Four iterations unrolled	Still fully unrolled; effectively unchanged

Performance

M5 Max, 16 Q / 2 KV heads, bf16, causal, 8192 x 8192; median of 12 runs:

Head dimension	Before	After	Change
128	5.70 ms / 48.2 TF	5.08 ms / 54.1 TF	+12% throughput
64	2.45 ms	2.44 ms	No material change

Scope and compatibility

• Changes one pragma in the existing wn=1 NAX QK loop.
• Does not change the head_dim == 256, wn=2 D-split path.
• Preserves the exact accumulation order.
• Composes cleanly with the head_dim == 256 PR; the two changes touch
  different regions of steel_attention_nax.h and can merge in either order.

Validation

• Outputs remain bitwise-identical because the loop-carried accumulation
  order is unchanged.
• Full local test_fast_sdpa.py result: 16 tests, 1 platform-dependent
  skip, otherwise passing.
• A three-way merge with the head_dim == 256 PR completes without conflict.

Checklist

• I have read the CONTRIBUTING document.
• I have run pre-commit run --all-files.
• Existing fast-SDPA tests pass with no numerical changes.
• No documentation changes are required for this compiler-scheduling change.