[cudax] Implement composite_mapping for groups

[davebayer]

This PR implements cudax::composite_mapping that allows to combine multiple mappings into one. This can allow the user to skip creating intermediate groups and allows much more flexibility. Users are not expected to create the composite_mapping instance directly, but rather by using the pipe operator |. For example:

cudax::group g{cuda::gpu_thread, cuda::this_warp{config}, cudax::group_by<4>{} | cudax::group_as<1, 3>{}, cudax::lane_synchronizer{});

Splits warp into 16 groups of counts 1 or 3.

Depends on #8755

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 43m 11s: Pass: 100%/54 | Total: 5h 59m | Max: 30m 59s | Hits: 97%/31953

See results here.