Refactor RMM resources to implement cuda::memory_resource concepts and deprecate and remove device_memory_resource and host_memory_resource base classes.

[harrism]

Part of #1443

Tasks that should not require prior work

Beta Give feedback

1. Rewrite rmm::device_scalar to take a cuda::mr::async_resource_ref instead of a device_memory_resource* #1446
   feature request +1
   
2. Rewrite rmm::exec_policy to take a cuda::mr::async_resource_ref instead of a device_memory_resource* #1448
   feature request +1
   
3. Deprecate rmm::polymorphic_allocator::resource and remove that functionality after deprecation #1453
   cpp feature request +2
4. Add rmm::device_check_resource_adaptor::get_upstream_resource
5. Remove duplicated tests that test both a device_memory_resource* and a cuda::mr::async_resource_ref #1450
   feature request +1
   
6. Rewrite multi_stream_allocation_benchmarks to take a cuda::mr::async_resource_ref #1481
   feature request +1
   
7. Add get_upstream_resource method to resource adaptors to eventually replace raw get_upstream #1455
   feature request +1
   

Tasks that would remove APIs that return a device_memory_resource

Beta Give feedback

1. Rewrite rmm::polymorphic_allocator to take a cuda::mr::async_resource_ref instead of a device_memory_resource* #1540
   +0
2. Rewrite rmm::device_check_resource_adaptor to take a cuda::mr::async_resource_ref instead of a device_memory_resource* #1541
   +0
3. Replace all internal usage of get_upstream with get_upstream_resource #1491
   3 - Ready for review breaking cpp feature request +4
4. Refactor Cython to use resource_ref instead of device_memory_resource* #1500
   0 of 1
   Python feature request +2
   

Final tasks that will have considerable downstream effects

Beta Give feedback

1. Rewrite rmm::per_device_resource to work with cuda::mr::async_resource_ref
2. Change rmm::get_current_device_resource to return a cuda::mr::async_resource_ref
3. Turn resource adaptors into alias templates #1457
   cpp feature request +2
   

[harrism]

For the ones with downstream effects we should build the new solution beside the existing one, then switch the downstream to the new, then deprecate the old, then in the following release remove the deprecated functionality.