Adds virtual shared memory helper and tests

[elstehle]

Description

This PR adds a virtual shared memory helper utility. This utility helps to work around the limitation of a CUB algorithm requiring more shared memory than is available per thread block (by default, 48K of static shared memory). This scenario usually only becomes relevant for large user types, e.g., once they exceed hundred(s) of bytes and more.

The utility has two mechanisms to work around this limitation:

1. The utility supports a fallback policy that adapts to this shared memory constraint. It achieves this by employing a smaller tile size, which means fewer threads per block and fewer items per thread. The memory requirements are intricately tied to the tile size and the data type the algorithm operates on. If the default policy surpasses the shared memory capacity while the fallback policy with a reduced tile size remains within these boundaries, the utility dynamically switches to the fallback policy. This ensures optimal resource usage without compromising performance.
2. If the algorithm with the fallback policy still exceeds the shared memory, we will provide the algorithm with virtual shared memory, which is memory backed by global memory.

Closes #553 and #1082

The plan is to put the new vsmem utility to use in a follow-up PR by integrating it into DeviceMergeSort for further evaluation. Once the API has proven to be robust, I'll add further documentation in the Developer Overview.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[alliepiper]

Some minor nits, but otherwise lgtm