Reapply "CUDA: fix race conditions FlashAttention kernels (ggml-org#13438)"

This reverts commit 6a199dd.

Author: Nexesenex

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -715,6 +715,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, WARP_SIZE);
         }
 
+        __syncthreads();
+
         // Write back combined meta data:
 #pragma unroll
         for (int imeta = 0; imeta < nmeta; ++imeta) {

ggml/src/ggml-cuda/fattn-vec-f16.cuh

@@ -181,6 +181,7 @@ static __global__ void flash_attn_vec_ext_f16(
     for (int j = 0; j < ncols; ++j) {
         KQ[j*D + tid] = -HALF_MAX_HALF;
     }
+    __syncthreads();
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};