Test every synchronous algorithm with a caching allocator to ensure that they correctly synchronize

In NVIDIA/thrust#1258, we ran into issues with radix sort in NVC++ because NVC++ uses a policy with a caching allocator, which means there is no implicit synchronization on freeing memory, so the algorithms need to actually have the explicit synchronization that we expect them to have. We should make sure we test every algorithm with a caching allocator to ensure this is the case.