fix: adjust missed double type

torch-ext/torch_binding.h

@@ -17,8 +17,7 @@ mha_fwd_kvcache_mla(
     const int64_t head_size_v,
     const torch::Tensor &seqlens_k,
     const torch::Tensor &block_table,
-    // TODO:should be float
-    const torch::kFloat softmax_scale,
+    const double softmax_scale,
     bool is_causal,
     const torch::Tensor &tile_scheduler_metadata,
     const torch::Tensor &num_splits