Thrust with tbb backend uses sequential versions for several algorithms

This extends what is reported in https://github.com/NVIDIA/cccl/issues/804

When compiled with tbb as the host backend, (at least) the following thrust algorithms use the _**sequential implementation**_ when `thruat::host`, or `thrust::tbb::par` execution policy is specified (in contradiction to the documentation), but use the correct tbb-based parallel implementation in overloads without the execution policy argument:

- thrust::copy_if
- thrust::remove_if
- thrust::transform_exclusive_scan
- thrurst::exclusive_scan (according to the previous report mentioned above)

I've been driven to create wrappers for all thrust algorithms I use in my library to redirect to the right version based on whether an algorithm belongs to this list of "buggy ones" or not. It is also fairly nontrivial to figure out which algorithms are buggy and until this is discovered for each buggy algorithm, considerable performance losses are accrued because of the incorrect behavior.