[EXPERIMENT] Use resource_ref for get/set_per_device_resource

[harrism]

Description

Components of this PR.

1. C++: Use rmm::device_async_resource_ref in get/set_per_device_resource(). (per_device_resource.hpp)
2. C++: Use the new API above in tests.
3. C++: Modify resource adaptor classes to store the upstream resource as a resource_ref rather than an MR pointer. Add a constructor that takes a resource_ref. (MR pointer constructor to be deprecated)
4. Python: Add minimal Cython bindings for device_async_resource_ref and cuda::stream_ref. Since we can't stack allocate device_async_resource_ref, add a get_ref() method on all Python DeviceMemoryResource classes that returns a shared_ptr[device_async_resource_ref]. This is dereferenced internally when passed to the bindings to the new cpp_set_per_device_resource. Refactor Cython for external resource classes.

This is currently an experiment.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[wence-]

I think the cython problems are not possible to work around right now :(, see cython/cython#6317

[harrism]

I think the cython problems are not possible to work around right now :(, see cython/cython#6317

I worked around them by using a shared_ptr[device_async_resource_ref] as discussed. I don't think the Python user ever needs to interact with this, so it's not too bad to deref it as needed.

[bdice]

I met with some folks from CCCL and cuda-python earlier this week to discuss memory resources. The new CCCL features in NVIDIA/cccl#2824 will change our strategy for owning memory resources. I will study this PR and will probably resurrect some parts of this in a new PR.