[FEA]: Add tests for large problem sizes in cub::DeviceMergeSort and fix integer overflow for problem sizes close to numeric_limits<OffsetT>::max()

[elstehle]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

CUB

Is your feature request related to a problem? Please describe.

After working on #1211 we should also add tests for large problem sizes (i.e., large num_items count).

Note, there are issues related to integer overflow when num_items is getting close to std::numeric_limits<OffsetT>::max(). In particular for 32-bit signed and unsigned types. So, a sub-task is to also fix potential integer overflows.

Describe the solution you'd like

• Adding tests for 32-bit offset types, with num_items being close to std::numeric_limits<OffsetT>::max().
• Adding tests for 64-bit offset types, with num_items being slightly above above std::numeric_limits<std::uint32_t>::max() to we are correctly computing offsets without wrapping around.
• Fix remaining integer overflows in offset computation when num_items is close to std::numeric_limits<OffsetT>::max()

Describe alternatives you've considered

No response

Additional context

No response