[Optimization] Specialize `localmap` and `localreduce`

[emmt]

localmap has some overheads due to the local buffer that must be allocated and filled. These overheads can be avoided in some specific cases:

• Specialize localmap so that localmap(mean,A,B) is the same as localmean(A,B); except that B can provides weight in localmean.
• Specialize localmap so that localmap(minimum,A,B) and localmap(maximum,A,B) is the same as erode(A,B) and dilate(A,B).
• Similarly localmap(sum,A,B) can be (if B is a simple box) optimized as localreduce(+,T,A,B) with suitable element type T.

[RoyiAvital]

Have you compared performance to other libraries? Specifically vs. Intel IPP?

[emmt]

No not yet. I do not have this commercial library.

[RoyiAvital]

It is available for free either as standalone or part of oneAPI:

https://www.intel.com/content/www/us/en/developer/tools/oneapi/ipp.html