WebGPU subgroups for attention and matmul shaders

[sauravpanda]

Context

WebGPU subgroups (shipping Chrome 144-145) enable warp-level SIMT operations. Benchmarks show up to 13x speedup, 26x with swizzle on some GPUs.

Proposal

• Add subgroup reductions to attention softmax (replace shared memory reduction)
• Use subgroup shuffle for matmul accumulation
• Feature-detect subgroup support, fallback to existing shaders
• Subgroup ballot for sparse attention patterns

References

• Chrome subgroups
• Chrome 145 added `subgroup_uniformity` extension