opencl: add Adreno xmem attention fast path

[happyyzy]

Summary

• add an Adreno-only OpenCL xmem attention route behind GGML_OPENCL_ADRENO_XMEM_ATTN
• add the xmem QK/PV kernels plus exact split-source variants used to preserve the fast Adreno compiler path
• route compatible ggml_flash_attn_ext cases to this path
• add test-opencl-adreno-attn as a focused correctness/perf reproducer

Scope

This PR is intentionally narrow. It only touches the ggml-opencl attention path:

• 3 new kernel files
• 1 runtime routing/integration file
• 1 benchmark test target

It does not change quantization, GEMM, CLML, or non-Adreno code paths.

Supported cases

• Adreno OpenCL only
• q: f32, k/v: f16 or f32, out: f32
• GQA
• causal and noncausal
• prefill and decode

Notes

The dedicated QK/PV split sources and their early compile order are intentional. On current Adreno drivers, compiling these xmem kernels later can produce a materially slower device binary even with identical source and build options.

Validation

Build validation on this branch:

• cmake -S . -B build-adreno-xmem -DGGML_OPENCL=ON -DGGML_OPENCL_EMBED_KERNELS=ON -DCMAKE_BUILD_TYPE=Release
• cmake --build build-adreno-xmem --target ggml-opencl test-opencl-adreno-attn -j
• ./build-adreno-xmem/bin/test-opencl-adreno-attn --help

Correctness examples on Adreno 830:

• noncausal + GQA: dq=128 dv=128 nq=256 nkv=512 n_head=8 n_head_kv=2 -> mae=6.7302e-05, max_abs=8.9369e-04, cos=0.999993535
• causal + GQA: dq=128 dv=128 nq=256 nkv=256 n_head=8 n_head_kv=2 -> mae=7.8839e-05, max_abs=9.92954e-04, cos=0.999999313
• decode: dq=128 dv=128 nq=1 nkv=512 n_head=8 n_head_kv=2 -> mae=7.4717e-05, max_abs=4.90941e-04, cos=0.999992445

Adreno 830 performance examples:

• H=30, L=4224, D=128, noncausal: xmem 188.606 ms / 1.453 TOPS; FA killed by system; nofuse OOM
• H=30, L=2048, D=128, noncausal: xmem 49.852 ms / 1.292 TOPS; FA 2015.568 ms / 0.03196 TOPS; nofuse 148.097 ms / 0.4350 TOPS
• H=30, L=2048, D=128, causal prefill: xmem 46.069 ms / 1.398 TOPS; FA 2150.036 ms / 0.02996 TOPS; nofuse 2155.344 ms / 0.02989 TOPS
• H=1, L=16384, D=512, noncausal: xmem 272.191 ms / 2.020 TOPS; nofuse 1084.935 ms / 0.5067 TOPS; FA hit map::at

Decode is not the target of this path and remains slower than the existing routes on small nq=1 cases:

• nq=1 nkv=512 dq=dv=128 n_head=8 n_head_kv=2: xmem 6.562 ms; FA 0.491 ms; nofuse 0.282 ms
• nq=1 nkv=16384 dq=dv=128 n_head=8 n_head_kv=2: xmem 12.350 ms; FA 8.276 ms; nofuse 3.535 ms