backends/cuda: use async malloc/free (pytorch#14976)

Found device synchronize in aoti_torch_delete_tensor_object via Linux
perf. This change appears to significantly improve self-reported latency
from voxtral_runner as found in
https://github.com/pytorch/executorch/blob/main/.github/workflows/cuda.yml#L111-L172:

Baseline:
Run latency (ms):
audio_encoder: 575.797
token_embedding: 14.571
text_decoder: 3095.356

With this PR:
Run latency (ms):
audio_encoder: 175.807
token_embedding: 8.799
text_decoder: 344.367

Author: swolchok

backends/cuda/runtime/shims/memory.cpp

@@ -225,7 +225,7 @@ AOTITorchError aoti_torch_empty_strided(
 
   if (device_type == static_cast<int32_t>(SupportedDevices::CUDA)) {
     ET_CUDA_CHECK_OR_RETURN_ERROR(
-        cudaMallocManaged(&ptr, static_cast<size_t>(nbytes)));
+        cudaMallocAsync(&ptr, static_cast<size_t>(nbytes), cudaStreamDefault));
   } else if (device_type == static_cast<int32_t>(SupportedDevices::CPU)) {
     // Ensure 16-byte alignment for CPU memory to match CUDA requirements
     int result = posix_memalign(&ptr, 16, nbytes);
@@ -328,11 +328,14 @@ AOTITorchError aoti_torch_delete_tensor_object(Tensor* tensor) {
           ET_CUDA_CHECK_OR_RETURN_ERROR(
               cudaPointerGetAttributes(&attributes, data_ptr));
 
-          if (attributes.type == cudaMemoryTypeManaged) {
-            // This is CUDA managed memory - free with proper synchronization
-            ET_CUDA_CHECK_OR_RETURN_ERROR(cudaDeviceSynchronize());
-            ET_CUDA_CHECK_OR_RETURN_ERROR(cudaFree(data_ptr));
+          if (attributes.type == cudaMemoryTypeDevice) {
+            ET_CUDA_CHECK_OR_RETURN_ERROR(
+                cudaFreeAsync(data_ptr, cudaStreamDefault));
           } else {
+            ET_CHECK_OR_RETURN_ERROR(
+                attributes.type != cudaMemoryTypeManaged,
+                Internal,
+                "Expected host memory but got managed!")
             // This is CPU memory - free immediately
             free(data_ptr);
             data_ptr = nullptr;