Keep scaled queries in F32 for causal GQA

[morluto]

Fixes #3.

Why

The causal GQA Metal kernel expanded a BF16 query to F32, applied the attention scale, then rounded that product back to BF16 before performing the QK contraction in F32:

Before: BF16 Q -> F32 scale -> BF16 round -> F32 QK contraction
After:  BF16 Q -> F32 scale --------------> F32 QK contraction

Because shared_query is already a threadgroup float array, the intermediate round does not save storage. It only discards precision before the scores and softmax are computed.

At head_dim=128, an A/B test of the real kernel over 64 deterministic trials and 2,097,152 outputs produced:

Metric versus explicit F32 attention	Before	After	Change
Maximum absolute error	0.0924071	0.00390625	95.77% lower
RMSE	0.00194536	0.000868307	55.37% lower
BF16 output mismatch rate	21.6468%	0.007486%	99.97% lower

The variants differed in 21.65% of output elements. Both were bit-for-bit deterministic.

What changed

• Keep bf16_to_f32(query) * scale in F32 through the QK contraction.
• Add a fixture-free regression test that exercises the production Metal kernel at head_dim=128 and compares it with an explicit F32 causal-GQA reference.
• Run the new test from the standard make test target.

The test uses numerical tolerances rather than device-specific hashes. A mutation check confirmed that it passes the fixed code and fails when the old BF16 re-round is restored:

Eight-trial regression	Maximum error	RMSE	BF16 mismatch rate
F32-scaled query	0.00390613	0.000864323	0.00839%
Old BF16 re-round	0.0924071	0.00197766	21.5992%

Performance

The real kernel was timed with 200 dispatches per sample, two series per variant, GPU command-buffer timestamps, and the first ramp-up sample in each series excluded.

Variant	Median dispatch time
BF16 re-round	21.274 us
F32 retained	20.491 us

The timing distributions overlap; the relevant conclusion is that no performance regression was observed.

Validation

Validated on Apple M5, macOS 26.3, arm64, Metal 4:

make -j"$(sysctl -n hw.logicalcpu)" h3_gqa_tests h3_audio_gpu_tests h3_tests
./h3_gqa_tests
./h3_audio_gpu_tests
./h3_tests
git diff --check

• h3_gqa_tests: passed
• h3_audio_gpu_tests: passed
• h3_tests: passed, 1,768 checks
• No Metal compilation or runtime errors

Scope

This kernel is used by the Qwen3-VL prompt encoder. The DiT attention path uses MPSGraph SDPA and is unchanged.

This PR is scoped to the directly reproduced kernel-level precision loss and its regression coverage.