Add DSA MLS sparse prefill dispatch

csrc/gfx93/prefill/sparse/dsa_mls/legacy/include/fwd/gfx938/pv_gemm_utils_mls_ds.h

+#pragma once // prepare for prefetch V in qk gemm
+#include "intrinsic_mls_ds.h"
+
+
+template<int kHeadDim, int kBlockM, int kBlockN, int kBlockK, int WARP_M, int WARP_N, typename Element, bool Is_even_MN>
+__forceinline__ __device__ void prefetch_v_to_lds_mls_ds(
+        vec4_uint v_ptr,
+        Element* v_lds,
+        int warp_id,
+        int seqlen_v_stride,
+        int max_seq_kv_offset=0) {
+
+    constexpr int ELEMENT_BYTES = sizeof(Element);
+    constexpr int WARP_NUM = kBlockM * kBlockN / (WARP_M * WARP_N);
+    constexpr int WARP_K   = 32;
+    constexpr int kHeadDim_OPT = (kHeadDim == 192) ? 128 : kHeadDim;
+
+    // MLS
+    // int n_loop = 0;
+    int n_load = 0;
+    vec4_uint v_srsrc;
+    if constexpr (kHeadDim == 128) {
+        *(uint64_t*)&v_srsrc = VA_LIMIT_BITS(*(uint64_t*)&v_ptr + (/*n_loop * WARP_K * seqlen_v_stride*/ + warp_id * 32) * ELEMENT_BYTES);
+    } else if constexpr (kHeadDim == 192) {
+        *(uint64_t*)&v_srsrc = VA_LIMIT_BITS(*(uint64_t*)&v_ptr + (n_load * WARP_NUM * 32 + warp_id * 32) * ELEMENT_BYTES);
+    }
+    v_srsrc[2] = seqlen_v_stride; // stride
+    if constexpr (true) {
+        int nm_filter = inline_min_max<0, 32>(0 * WARP_K + 32 - max_seq_kv_offset);
+        v_srsrc[3] = max_seq_kv_offset % kBlockK == 0 ? 0: nm_filter << 8;
+        v_srsrc[3] += 0x20000;
+    }
+    int lds_stage_id = 0;
+    int lds_offset = (lds_stage_id * WARP_K * kHeadDim_OPT + warp_id * 32 * 32) * ELEMENT_BYTES;
+
+    flash::wait_all_warp_arrived(); // 防止写 v lds 和读 k lds 冲突, qk 可能有的 warp 没结束
+    inline_matrix_load_32x32_b16_lds<0, 1>(v_lds, v_srsrc, lds_offset, 0);
+    __builtin_amdgcn_sched_barrier(0);
+}
+