b6794

vulkan: Implement topk_moe fused shader, ported from CUDA (#16641)

This is similar to the CUDA shader from #16130, but doesn't use shared memory
and handles different subgroup sizes.