simplify mapreduce_impl

[vtjnash]

this is largely now possible since Functors are not needed for performance
fixes #16185

[KristofferC]

I don't think we should close #16185 until the reduced example I posted there is fixed. Alternatively a new issue could be made for that example.

[JeffBezanson]

The reduced example in that issue might not get fixed. The extra code allows dynamic dispatch to a version of g specialized for f. The method in the example is probably not performance-critical, since all it does is call another function. This is part of the cost of faster higher-order functions.

[StefanKarpinski]

From triage call:

• addresses the performance issue in performance regression in sum(a) #16185
• by changing the association order of mapreduce
• does not address the dynamic dispatch slowdown that causes performance regression in sum(a) #16185

[StefanKarpinski]

TODO: there should be a note in NEWS about the changed reduction order.

[JeffBezanson]

We should definitely keep both pairwise summation and the 0.4 level of performance. If necessary we should manually specialize sum, or use a reduce_impl that has the mapping function removed.

[tkelman]

Apparently DataArrays was using sum_pairwise_blocksize

[vtjnash]

it can define sum_pairwise_blocksize(T) = pairwise_blocksize(T, +) for compatibility

[nalimilan]

That's what I did in JuliaStats/DataArrays.jl#192. If that was the only use of it outside of Base, I guess that's OK.

[tkelman]

NullableArrays was also using it.