Add tests for large number of items to DeviceSelect

[elstehle]

In PR #1379, we consolidated thrust::copy_if to use cub::DeviceSelect. Along with that change, we made thrust to dispatch using offset types using i64 if the given number of items was exceeding INT_MAX.

Testing for large number of items in thrust::copy_if has been limited to some sanity tests.

We want to have more elaborate test coverage by adding tests for large number of items to the CUB tests of cub::DeviceSelect, since we have better testing infrastructure in CUB.

However, until we agree on a scheme for offset type handling in #50, cub::DeviceSelect still only uses hard-coded int offset types. To be able to still test DeviceSelect for large number of items, we want to directly use the Dispatch layer, which allows instantiating the algorithm for wider offset types.