Speed up some use cases of ArrayPartition

[mateuszbaran]

This is the first batch of changes related to #160 . Most changes are related to the fact that map is much more friendly to Julia compiler than broadcasting so I've replaced a bunch of unnecessary broadcasts with map. The other thing is that the loop in assignment broadcast is very problematic -- ideally compiler should be able to constant-propagate N and unroll the loop but it seems that doing so breaks some other optimizations. This change helps a bit but it's still not enough.

[mateuszbaran]

I've improved that copyto! method and now my example from #160 is about 100x faster than before 🎉 :

julia> @benchmark f!($c, $a, $b)
BenchmarkTools.Trial: 10000 samples with 999 evaluations.
 Range (min … max):  7.885 ns … 34.613 ns  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     7.972 ns              ┊ GC (median):    0.00%
 Time  (mean ± σ):   7.990 ns ±  0.406 ns  ┊ GC (mean ± σ):  0.00% ± 0.00%

                                  ▂▂█▁▄▄                      
  ▂▂▂▂▁▁▂▁▂▂▁▂▂▂▂▂▂▃▃▃▃▃▃▄▄▅▄▅▄▄▇▆██████▇▇▄▅▅▄▇▆▇▄▄▃▃▂▂▃▂▃▃▃ ▃
  7.88 ns        Histogram: frequency by time         830 ns <

 Memory estimate: 0 bytes, allocs estimate: 0.

[ChrisRackauckas]

oh yes, just mapping over the tuple is a lot better, great idea.