But are they an abstraction that will be useful in the future? Worth it for the language?
Normally it's good to have pleasant abstractions and generic functions. But execution policies are there for speed and speed alone: if they don't make things faster, they've failed. So in this case it's worth tolerating a more awkward API to achieve better performance.
The C++17 parallelism interface is more or less to declare "this loop can be parallelized," which does not provide enough information to the scheduler. You need to express properties of the algorithm, such as its optimal chunk size. (In the future perhaps those properties will enter into the execution policy as e.g. constructor parameters.)
Second, passing in callbacks which receive one or two elements at a time is too granular. You are at the mercy of the compiler to optimize these. A better approach would be for the callbacks to receive a pair of iterators and allow the user to drive the loop. This makes it easier to avoid accidental indirections, failed alias analysis or inlining, etc.
Intel's TBB will continue to be more useful here.