Implement unified ETLS without TBB and use it for all host backends

[danhoeflinger]

tbb/enumerable_thread_specific.h includes an indirect inclusion of windows.h on windows.
This is undesirable, and therefore we would like to avoid it. Without enumerable_thread_specific to depend upon in TBB, we can instead unify all implementations into a shared one which uses a pair of helpers:
__get_thread_num()
and
__get_num_threads()
which are passed to the shared implementation from the parallel backends as template arguments during creation in the make function.

We implement a make function for each backend to properly pass in the appropriate helper functions for the parallel backend. We no longer need to bury the __enumerable_thread_local_storage in a __details namespace, since all backends could in theory directly use the constructor if they fill out all the appropriate template arguments. Variadic Args... are present in the type itself for all backends including TBB now. However, make functions are beneficial to keep verbosity out of the local usage of __enumerable_thread_local_storage.

---

Sadly, I don't see a way to automatically pick up the appropriate helper routines from the parallel backend from within include/oneapi/dpl/pstl/parallel_backend_utils.h because of a circular dependency without adding another utils header.

We could consider adding another header include/oneapi/dpl/pstl/parallel_unified_utils.h which depends on the parallel backend rather than vice-versa. This could allow us to depend upon a generic __par_backend:: call for the two helpers. We could also unify all make functions into this header. The downside to this would be that we sacrifice the ability to use ETLS from the backends themselves, which we dont do now, but we may want to do in the future.

[SergeyKopienko]

@danhoeflinger, I haven't any additional questions or comments from my side.

[mmichel11]

The approach here LGTM.

[akukanov]

Looks good.

[mmichel11]

LGTM

[danhoeflinger]

I had to revert CRTP implementation due to an issue on CL2017. Back to passing callables details from the individual parallel backends. Otherwise, nothing has changed since approval. CI run is going now.

[mmichel11]

Reapproving