opencl: cluster-parallel decode FA for Adreno GPUs

[wanghqc]

Overview

• Motivation of this work: Decode flash attention on Adreno was latency-bound (20-42% of the bandwidth floor): one workgroup per Q-head leaves most of the GPU idle at ne1==1.
• This work is to add cluster-parallel decode FA kernels (f16, q4_0, q8_0 KV) that split each head's KV range across a cluster of workgroups with a cross-workgroup reduction.
• Cluster width is a register-spill gate, and the optimal value depends on GPU and KV types.
• This can be override by using the env variable GGML_OPENCL_FA_CL_C= {8,16,32}.
• It is turned on by default for X1 and X2 GPUs, and can be turned off by setting env var GGML_OPENCL_FA_C8 = 0;
• For other GPUs set GGML_OPENCL_FA_C8 = 1 to turn on.

Additional information

• Performance gain: Decode-at-depth (Adreno X2-90, Qwen3-30B-A3B): +47%/+62% tg at d8k/16k vs the earlier PR 25366, reaching 74% of the bandwidth floor; up to +74-93% at 16k across models.
• May work with Intel GPUs.

Requirements

Tested with Windows on Snapdragon (WoS) Adreno GPUs, X1 and X2.

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

[max-krasnyansky]

Fancy! :)