RFC: add default dims to eachslice

[mcabbott]

I wonder if eachslice should have a default of the last dimension.

In python you can do for mat in batch to iterate a 3D array by getting its slices... the ones which make sense for row-major arrays. With this, something like for mat in eachslice(batch) would iterate the slices which make sense for column-major arrays.

julia> sizes(A) = (inner = size(first(A)), outer = size(A));

julia> eachslice(rand(2,3,5,7)) |> sizes  # previously no default
(inner = (2, 3, 5), outer = (7,))

julia> stack(eachslice(rand(2,3,5,7))) |> size  # matches this
(2, 3, 5, 7)

Edit: closes #54626

Extend to eachcol? (Edit: removed for now)

Less obviously, we could also extend eachcol to always give a slice along the first dimension. Then for col in eachcol(batch) would return dense 1D views for any higher-dim Array.

julia> eachcol(rand(2,3,5)) |> sizes  # always 1st dim inner
(inner = (2,), outer = (3, 5))

julia> eachcol(rand(2,3)) |> sizes  # as before
(inner = (2,), outer = (3,))

julia> eachcol(rand(2,)) |> sizes  # as before, involves a reshape
(inner = (2,), outer = (1,))

For symmetry we should then extend eachrow; note that on vectors it currently picks the 2nd (trivial) dimension, so I think that's what it would have to do always:

julia> eachrow(rand(2,3,5)) |> sizes  # always 2st dim inner
(inner = (3,), outer = (2, 5))

julia> eachrow(rand(2,3)) |> sizes  # as before
(inner = (3,), outer = (2,))

julia> eachrow(rand(2)) |> sizes  # as before,  involves a reshape
(inner = (1,), outer = (2,))