Let BlockDropout reuse LDS with V

include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp

@@ -533,9 +533,8 @@ struct BlockFmhaPipelineQRKSVSAsync
 
             if constexpr(kHasDropout)
             {
-                auto randval_ptr = reinterpret_cast<char*>(smem_ptr) +
-                                   Policy::template GetSmemSizeK<Problem>() +
-                                   Policy::template GetSmemSizeV<Problem>();
+                auto randval_ptr =
+                    reinterpret_cast<char*>(smem_ptr) + Policy::template GetSmemSizeK<Problem>();
                 dropout.template Run<decltype(gemm_0), SMPLComputeDataType, RandValOutputDataType>(
                     smem_ptr, seqlen_k_start + i_total_loops * kN0, p_compute, randval_dram_window);
             }

include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async_default_policy.hpp

@@ -177,8 +177,10 @@ struct BlockFmhaPipelineQRKSVSAsyncDefaultPolicy
     CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize()
     {
         // assume Q can reuse the shared memory with K or V
-        return max(GetSmemSizeQ<Problem>(), GetSmemSizeK<Problem>() + GetSmemSizeV<Problem>()) +
-               GetSmemSizeDropout<Problem>(0);
+        // assume Dropout can reuse the shared memory with V
+        return max(GetSmemSizeQ<Problem>(),
+                   GetSmemSizeK<Problem>() +
+                       max(GetSmemSizeV<Problem>(), GetSmemSizeDropout<Problem>(0)));
     }
 };