Asynchronous memory allocation for GPUs

Syncronous memory allocation can have a lot of overhead on GPUs. This was noticed when doing tests for multi-block filters. This is supported in cuda versions > 11.3

This is a redo of !2759 (closed). Much has changed since that MR so this will be the replacement.

vtkm/cont/testing/UnitTestArrayHandle.cxx

@@ -16,10 +16,6 @@
 #include <vtkm/cont/ArrayHandleExtractComponent.h>
 #include <vtkm/cont/testing/Testing.h>
 
-#include <vtkm/cont/DeviceAdapterTag.h>
-#include <vtkm/cont/RuntimeDeviceInformation.h>
-#include <vtkm/cont/cuda/internal/CudaAllocator.h>
-
 #include <algorithm>
 #include <vector>
 
@@ -552,7 +548,7 @@ struct VerifyFill
   }
 };
 
-VTKM_CONT void RunTests()
+VTKM_CONT void Run()
 {
   vtkm::testing::Testing::TryTypes(VerifyEmptyArrays{});
   vtkm::testing::Testing::TryTypes(VerifyUserOwnedMemory{});
@@ -565,20 +561,6 @@ VTKM_CONT void RunTests()
   vtkm::testing::Testing::TryTypes(VerifyFill{});
 }
 
-VTKM_CONT void Run()
-{
-#ifdef VTKM_CUDA
-  //For cuda, run tests with and without managed memory.
-  //When managed memory is off, async cuda malloc will be used.
-  vtkm::cont::cuda::internal::CudaAllocator::ForceManagedMemoryOff();
-  RunTests();
-  vtkm::cont::cuda::internal::CudaAllocator::ForceManagedMemoryOn();
-  RunTests();
-#else
-  RunTests();
-#endif
-}
-
 } // anonymous namespace
 
 int UnitTestArrayHandle(int argc, char* argv[])