perf: extend MMA decode to arbitrary GQA ratio, add launch bounds, vectorize combine

- Multi-pass MMA: encode pass in grid blockIdx.x, compute q_head0/G in-kernel
- Fixes crash for G>32 (previously block(32,G) exceeded 1024 threads)
- Fixes alloc_split_partials using uninitialized num_splits (MAX_SPLITS=32)
- __launch_bounds__ on all MMA and prefill kernels for better register allocation
- 4x vectorized combine kernel (4 head_dim per thread)
- uint4 vectorized K loads in scalar decode kernels
- cp.async .L2::128B cache hint for K/V tile streaming
- Extract warp_reduce_sum, bf16, MAX_SPLITS to attn_warp_utils.cuh

csrc/kernels/attn_warp_utils.cuh

@@ -0,0 +1,13 @@
+#pragma once
+#include <cuda_bf16.h>
+
+using bf16 = __nv_bfloat16;
+
+static constexpr int MAX_SPLITS = 32;
+
+__device__ inline float warp_reduce_sum(float val) {
+    #pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1)
+        val += __shfl_xor_sync(0xFFFFFFFF, val, offset);
+    return val;
+}