WIP: Generalize @simd to multi-dimensional ranges.

[ArchRobison]

This is work in progress to address #9858. I'll need help from @timholy to write the three necessary simd_... overloads for CartesianRange.

[timholy]

Very nice. I will submit a PR against your PR. The exciting bit: given the function definitions in #9858,

A = rand(Float32, 10^4, 10^4);

julia> @time f3(A)
elapsed time: 0.225945334 seconds (96 bytes allocated)
4.9999244533296704e7

julia> @time f3_broken(A)
elapsed time: 0.078786508 seconds (96 bytes allocated)
4.9999244533296704e7

julia> @time f1(A)
elapsed time: 0.084660836 seconds (96 bytes allocated)
4.9999244533296704e7

julia> @time f2(A)
elapsed time: 0.131078207 seconds (96 bytes allocated)
4.9999244533296704e7

So with @simd, we're back to being just as fast as manually-written nested loops. (Presumably this is true only for arrays with a sizable first dimension, but still...)

[timholy]

Incorporated via #9876. Thanks, @ArchRobison, for tackling this so quickly!

[timholy]

This gave me a 3.5-fold speed improvement in a computation I care a lot about, just by adding @simd in front of two lines. Thanks!