Increase CUB/Thrust benchmark problem sizes

The benchmarks for CUB and Thrust usually pick the problem sizes 2^16, 2^20, 2^24 and 2^28. The largest size using `int32` results in 1GiB of memory used (per input), which is not a large workload on recent data center grade GPUs. On B200 for the new warpspeed scan implementation it was barely large enough to hide the constant overheads of the kernel and reach SOL.

We should strongly consider adding a larger benchmark size, e.g. 2^32, which is only executed for 64-bit offset types. Since this may fail to allocate on some GPUs of interest, we should then gracefully skip the benchmark if not enough memory is available.