opencl: general flash attention decode performance optimizations

[wanghqc]

Overview

• This PR has a comprehensive performance optimization on token generation (decode) for Adreno GPUs, i.e., when n_q == 1. In particular, we target token generation performance with the long context window of 8k, 16k and beyond for agentic use cases.

• Non-Adreno devices keep the original path.

Additional information

Details of the features in this PR:

1. Vec decode kernels for f16/q8_0/q4_0 KV (DV-split, subgroup-reduced dots), with a flash-decoding KV split for deep contexts
2. KV-head-coalesced multi-query decode kernels for GQA models (MQ_GQA 4 and 8);
3. Fix an OOO issue with Adreno compiler by compiling the MQ_GQA=8 kernels in a separate minimal program
4. Multi-output and GQA-coalesced mul_mat kernels for the KQ/KQV matmuls when flash attention is off, with option of using image1d_buffer_t
5. DK=DV=512 (gemma-4 global-attention layers): prefill runs on the GPU via split programs; decode is offloading to CPU for now due to CPU's performance edge for this case (effective GPU BW < CPU BW)
6. f16-KV decode split granularity: 512 KV/split (from 2048); For quantized KV keeps 2048
7. A few correctness fixes along these paths: finite running-max init, pooled buffers for the quantized-K dequant path, build retry on transient compiler OOM, trans-weight restore for broadcast quantized mul_mat, etc.

Requirements

Fully tested with Adreno X1 and X2 Windows on Snapdragon (WoS) devices.

• I have read and agree with the contributing guidelines Yes.
• AI usage disclosure: Yes, for testing and quick prototyping.