hexagon: flash attention rework (optimizations, accuracy improvements, etc)

[max-krasnyansky]

Overview

The main focus here is to update FLASH_ATTN_EXT in ggml-hexagon similar to what I did in #24954

• Combining HMX and HVX into htp/flash-atten-ops.c and removing duplication
• Updating op tracing instrumentation (added a few new events)
• Moving inner kerners into htp/{hvx,hmx}-fa-kernels.h
• Moving kernel-params computation into the host (cached as part of the graph caching)
• Improving pipelining and most of the kernels
• Updated softmax to use FP32 accumulators.
  Similar to hexagon: store HMX flash-attention softmax accumulators in FP32 #24389

The PR also includes some MUL_MAT updates as a followup to #24954.
As I was looking at Flash Attention op-traces I noticed a few more things to improve in MUL_MAT.
Technically, I could split that up but I did extensive testing of the combined update.

Additional information

Tested on all supported devices.
Newer devices show higher gains but there are steady improvements across the board

Details

## gemma-4-E2B_q4_0

Ventuno-Q
   prompt eval time = 1466.41 ms / 786 tokens (  1.87 ms per token,  536.00 tokens per second) (vs 507.80 master)
          eval time = 4114.31 ms /  63 runs   ( 65.31 ms per token,   15.31 tokens per second) (vs  15.04 master)

S24U
   prompt eval time =  848.53 ms / 741 tokens (  1.15 ms per token,  873.28 tokens per second) (vs 815.58 master)
          eval time = 2816.80 ms /  63 runs   ( 44.71 ms per token,   22.37 tokens per second) (vs  19.29 master)

S25+
   prompt eval time =  497.93 ms / 741 tokens (  0.67 ms per token, 1488.15 tokens per second) (vs 1201.47 master)
          eval time = 2324.59 ms /  63 runs   ( 36.90 ms per token,   27.10 tokens per second) (vs   25.24 master)

S26+
   prompt eval time =  403.29 ms / 741 tokens (  0.54 ms per token, 1837.38 tokens per second) (vs 1646.45 master)
          eval time = 1983.48 ms /  63 runs   ( 31.48 ms per token,   31.76 tokens per second) (v    29.45 master)

X2-Elite
   prompt eval time =  376.89 ms / 741 tokens (  0.51 ms per token, 1966.10 tokens per second) (vs 1688.41 master)
          eval time = 1465.07 ms /  63 runs   ( 23.26 ms per token,   43.00 tokens per second) (vs   34.15 master)

## Qwen3.5-2B-Q4_0.gguf

S24U
   prompt eval time = 1249.01 ms / 742 tokens (  1.68 ms per token,  594.07 tokens per second) (vs 581.78 master)
          eval time = 2847.99 ms /  63 runs   ( 45.21 ms per token,   22.12 tokens per second) (vs  21.59 master)

S25+
   prompt eval time =  753.36 ms / 742 tokens (  1.02 ms per token,  984.92 tokens per second) (vs 969.03 master)
          eval time = 2232.04 ms /  63 runs   ( 35.43 ms per token,   28.23 tokens per second) (vs  27.29 master)

S26+
   prompt eval time =  570.42 ms / 742 tokens (  0.77 ms per token, 1300.80 tokens per second) (vs 1282.19 master)
          eval time = 2015.43 ms /  63 runs   ( 31.99 ms per token,   31.26 tokens per second) (vs   30.79 master)

X2-Elite
   prompt eval time =  568.55 ms / 742 tokens (  0.77 ms per token, 1305.07 tokens per second) (vs 1258.02 master)
          eval time = 1779.24 ms /  63 runs   ( 28.24 ms per token,   35.41 tokens per second) (vs   37.17 master)

## Llama-3.2-1B

S26+
   prompt eval time =  190.18 ms / 766 tokens (  0.25 ms per token, 4027.72 tokens per second) (vs 2855.28 master)
          eval time = 1161.91 ms /  63 runs   ( 18.44 ms per token,   54.22 tokens per second) (vs   42.41 master)

## Llama-3.2-3B

S26+
   prompt eval time =  421.11 ms / 766 tokens (  0.55 ms per token, 1819.01 tokens per second) (vs 1431.30 master)
          eval time = 2519.05 ms /  63 runs   ( 39.98 ms per token,   25.01 tokens per second) (vs   21.45 master)

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, lots of help from Antigravity to explore different ideas, and refactoring.