Reductions of Transpose, Adjoint, PermutedDimsArray

[mcabbott]

Closes #33029.

It also adds some methods to ∘ such that adjoint ∘ identity ∘ adjoint === identity, since it will tend to produce such combinations, and something else may wish to dispatch on them. But it would work without this, I can separate it if that's preferable. (Perhaps we should also complete the group with conj ∘ transpose = adjoint etc.) Now removed, in favour of an explicit _sandwich function.

The reduction functions which this overloads for PermutedDimsArray were defined after its file is loaded. I moved this file later, but could equally well move these methods to another file.

[mcabbott]

Bump. Perhaps paging @tkf, since you were involved in #33029

[mcabbott]

This doesn't seem speed up maximum:

julia> X = randn(5000, 200);

julia> @btime maximum($X);
  636.708 μs (0 allocations: 0 bytes)

julia> @btime maximum(transpose($X));
  4.643 ms (0 allocations: 0 bytes)

julia> @btime maximum($X; dims=1);
  744.458 μs (4 allocations: 1.88 KiB)

julia> @btime maximum(transpose($X); dims=2);
  4.644 ms (14 allocations: 2.06 KiB)

although sum works as expected:

julia> @btime sum($X);
  200.250 μs (0 allocations: 0 bytes)

julia> @btime sum(transpose($X));
  200.250 μs (0 allocations: 0 bytes)