Separate the datapar algorithms

[hkaiser]

Currently, our parallel algorithms support being used with the datapar execution policy. This is a remnant of an older standardization proposal. This implementation is currently tightly entangled with the implementation of the parallel base algorithms.

We should do two things:

• separate the datapar implementations and expose them through separate algorithm specializations (based on the tag_invoke customization point mechanism we have implemented)
• adapt the implementation to support (and rely on) the data-parallel Types introduced by N4755, section 9, this implies removing datapar as it is today,

This work could also go hand-in-hand with #2271.

[toktarev]

I've started working on this issue

[hkaiser]

@toktarev Please coordinate with @srinivasyadav18 as he is looking into related tasks.

[toktarev]

Do we really need to "separate the datapar implementations" if we are going to "removing datapar as it is today" ?