hexagon: store HMX flash-attention softmax accumulators in FP32

[njsyw1997]

Overview

Store the online-softmax cross-block accumulators of the HMX flash-attention prefill kernel in FP32 instead of FP16. m (row max), l (row sum), and p_rowsum move to FP32; s_rowmax stays FP16 (lossless — it's a max of fp16 values).

Additional information

Q/K/V are FP16 on both the HMX path and the Metal reference. However, the HMX kernel also keeps the running softmax statistics (m/l) and the per-block p_rowsum in FP16 while these intermediate results are stored in fp32 in Metal. Since these values are re-quantized to FP16 for each KV block, the error accumulates over time.

PPL does not directly capture this shift, but in some real long-context examples, we do observe a divergence from the CPU backend, while Metal remains aligned.

This PR will slow down prefill speed by roughly 5%.

Potential bug

It works before #23796. After that PR, it may randomly produce incorrect results in some builds on Qwen3 4b Q4_0. Similar to the previous issue, a good build consistently produces correct results, while a bad build consistently produces incorrect results.

Adding #define FARF_HIGH 1 to print log fixes the problem, which suggests that the bug is likely in one of the synchronization procedures.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes. Mainly for analyzing the commit history and constructing tests. All the code is reviewed by me.

[njsyw1997]

@max-krasnyansky
Hi Max, can you have a look?