Support named dims for dims keyword argument

[bencottier]

To use a consistent convention for dims (#18), we currently have to "invert" the value passed into the dims keyword argument to the array-based apply!, so that eachslice gives the expected result. For example dims=2 becomes dims=1 and dims=3 becomes dims=[1, 2]. Basically we can use setdiff(1:ndims(array), dims) to accomplish that.

The solution isn't so straightforward for named dims, because the method to get the list of dim names is not universal. KeyedArray supports dimnames(A) but AxisArray does not, instead using axisnames. One workaround is dimnames(wrapdims(::AxisArray)) but that requires adding AxisKeys as a dependency.

Some have suggested that support for AxisArrays should be dropped entirely, or included in an auxiliary package - then we could rely on dimnames.

[glennmoy]

"Not having named dims limits the usefulness of using a KeyedArray as opposed to just an AbstractArray"

[nicoleepp]

I personally vote that adding the dependency on some of the underlying packages might be worth having named dims for

[bencottier]

Are named dims actually prevented anywhere now? I think it may only be keys for inds now. I just checked named dims works for MeanStdScaling on KeyedArray.

[bencottier]

I personally vote that adding the dependency on some of the underlying packages might be worth having named dims for

I don't have a good sense of the cost of adding the dependency. On the face of it, it seems worthwhile if we can't find any other non-hack way.

[glennmoy]

Are named dims actually prevented anywhere now? I think it may only be keys for inds now. I just checked named dims works for MeanStdScaling on KeyedArray.

They are not in fact. We've added explicit tests in #62