This is special for arrays of shape 1 x n or n x 1; add iterator and
iterator sum (fold) benchmarks.

Implement axis merging - this preserves order of elements in the
iteration but might simplify iteration. For example, in a contiguous
matrix, a shape like [3, 4] can be merged into [1, 12].

Also allow arbitrary order optimization - we then try to iterate in
memory order by sorting all axes, currently.

This one complains about if let Some(_) = option; which I still think is
just as fine as if option.is_some(); `if let` is a nice Rust feature
that works the same way on all enums, no reason to prefer
option-specific methods like `is_some`.