[FEA] Make rmm::prefetch fault tolerant

[bdice]

Following up from rapidsai/cudf#16551 (comment)

Two problems:

• Currently RMM does not document the requirement that a prefetch_resource_adaptor must wrap a managed memory resource.
• Additionally, on systems where managed memory is not supported, a cudaErrorInvalidDevice occurs.

We should make prefetching a transparent fault-tolerant API, with no hard errors. This makes it easier to use prefetching as a "hint" in an application without having to conditionally check for managed memory support from the current memory resource before calling the prefetch API. In cuDF, for instance, we would like to enable prefetching as a performance optimization in many locations -- and do nothing if the memory resource isn't compatible.