use ObservationsDims.jl to extend fit and predict to be observation aware

[oxinabox]

I suggest the way to do this is that fit and predict should stay as they.
They are a consistent API with observations always a columns.

Further, ObservationDIms does not support a different obs_arrangement for different methods.
Only for different functions.
So can only declare it once for fit and predict, not once per model.

So we have another function e.g:

function smart_fit(template, outputs, inputs, weights=nothing; obsdim=nothing, kwargs...)
    outputs = organise_obs(MatrixColsOfObs(), outputs; obsdim=obsdim)
    inputs = organise_obs(MatrixColsOfObs(), inputs; obsdim=obsdim)
    if weights === nothing
        weights = uweights(size(outputs, 2))
    end
    return @mock fit(template, outputs, inputs, weights; kwargs...)
end

function smart_predict(model, inputs; obsdim=nothing, kwargs...)
    new_inputs = organise_obs(MatrixColsOfObs(), inputs; obsdim=obsdim)
    permute = xor(new_inputs isa PermutedDimsArray, inputs isa PermutedDimsArray)
    outputs = @mock predict(model, new_inputs, weights; kwargs...)
    if permute
        outputs = PermutedDimsArray(outputs, (2, 1))
    end
    return outputs
end