Accept non-AbstractArrays in 2-arg axes/size

[jishnub]

With this change, any type that implements axes(A) gets axes(A, d) for free, and similarly for size. These would make it easier to accept array-like types such as Broadcasted in functions that require indexing.

One caveat is that the default assumption here is that for d > ndims(A), we return a single-element axis OneTo similar to what we do for AbstractArrays. This may not always be meaningful for arbitrary types, and indexing with trailing unit indices might not be supported. Such types currently implement axes(A, d) for themselves and should continue doing so.

[adienes]

should ndims then have a fallback length(axes(A)) ? and size one map(length, size(A))? although that might conflict with the existing

function axes(A)
    @inline
    map(unchecked_oneto, size(A))
end

[jishnub]

Good point, perhaps we may require Base.IteratorSize(A) to be specialized to return an Base.HasShape{N}, and infer ndims from this? I imagine most iterators wouldn't want to opt into the multidimensional array-like behavior that comes with ndims being automatically inferred from size/axes.

[adienes]

seems fair to me. as long as it is not the other way around (setting IteratorSize based on axes or size ...)

[nsajko]

some overlap with my PR #59465

[nsajko]

With this change, most two-argument methods of size/axes should perhaps be deleted. In this PR or in followups. However some of that would overlap with my open PR #59442 (which depends on either PR #59465 or this very PR getting merged before it can be mergeable).

[jishnub]

@nanosoldier runbenchmarks(ALL, vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.