Fix memory OOM issue (#479)

This pull request includes changes to improve memory management in
GPU-related functions by ensuring proper release of memory handles.
Fix #470

Author: Binyang2014

src/gpu_utils.cc

@@ -144,6 +144,7 @@ void gpuFreePhysical(void* ptr) {
   CUmemGenericAllocationHandle handle;
   size_t size = 0;
   MSCCLPP_CUTHROW(cuMemRetainAllocationHandle(&handle, ptr));
+  MSCCLPP_CUTHROW(cuMemRelease(handle));
   MSCCLPP_CUTHROW(cuMemGetAddressRange(NULL, &size, (CUdeviceptr)ptr));
   MSCCLPP_CUTHROW(cuMemUnmap((CUdeviceptr)ptr, size));
   MSCCLPP_CUTHROW(cuMemRelease(handle));

src/registered_memory.cc

@@ -75,6 +75,7 @@ RegisteredMemory::Impl::Impl(void* data, size_t size, TransportFlags transports,
       MSCCLPP_CUTHROW(
           cuMemExportToShareableHandle(transportInfo.shareableHandle, handle, getNvlsCompatibleMemHandleType(), 0));
       transportInfo.offsetFromBase = (char*)data - (char*)baseDataPtr;
+      MSCCLPP_CUTHROW(cuMemRelease(handle));
     } else {
       cudaIpcMemHandle_t handle;
       MSCCLPP_CUDATHROW(cudaIpcGetMemHandle(&handle, baseDataPtr));