Uvm backend

[dllehr-amd]

Add CUDAMallocManagedAllocator Backend

With the new CUDAAllocator class, we have created a new
CUDAMallocManagedAllocator, which will handle allocator requests from both
cpu and cuda device types when the backend is enabled
You can enable the backend using PYTORCH_CUDA_ALLOC_CONF=backend:cudaMallocManaged
And view inside PyTorch using torch.cuda.get_allocator_backend()

This allocator is initially rudimentary as the performance implications of a
managed allocator are still being worked out. However, the goal is to be able to
swap out the backend when running without any code change required.