They were previously pub, just because we didn't have the pub(crate)
feature yet.

Using split tests performance of the Zip in parallelization, so that we
can see if there are benefits to splitting arrays better.

Using the index shows more directly the overhead of indexed zip

Support both unroll over c- and f-layout preferred axis in Zip inner loop
(the fallback when inputs are not all contiguous and same layout).

Keep a tendency score when building the Zip, so that we know if the
inputs are tending to be c- or f- layout.

This improves performance on the just added zip_indexed_ff benchmark, so
that it seems to match its (already fast) cc counterpart.