Use the new CUB tuning APIs internally

[bernhardmgruber]

After the tuning API has been (internally) implemented for all CUB algorithms, we can start using the new tuning APIs already internally. This should mostly affect benchmarks. This should give us time to iron out any issued with the design:

• Ensure all CUB algorithms correctly extract an eventual tuning from their environment. This should include adding a test for each CUB algorithm overload specifying a custom tuning.
• Replace all internal dispatch layer access by public CUB API calls including the new tuning API. This mostly affects benchmarks.