Since pytorch/pytorch#91398, the signature of
the pluggable allocate and deallocate functions must accept the device
id. The current version only accepts a device id for allocate, which
means that when using a stream ordered allocator with devices other
than device zero, we pass an invalid stream into the deallocation
function. To fix this, adapt the signature to match the one pytorch
expects.

Now, since we have the device available during allocation and
deallocation, we would like to use that device to obtain the
appropriate memory resource.

Unfortunately, since RMM's cuda_device_id does not have a nullary
constructor, we can't use it in Cython without some hacky workarounds.

However, since we don't actually need to build a Python module, but
rather just a single shared library that offers two extern "C"
functions, let's just write our allocator hooks directly in C++.

- Closes rapidsai#1405

PyTorch does not guarantee that plug-in allocator functions are called
with the active device matching the device on which the
allocation/deallocation is requested. Hence, we must explicitly handle
that here by selecting the appropriate device.