Adds support for using Base.mapreducedim! with AbstractField and AbstractOperations

[glwagner]

This PR adds support for in-place reduction operations where the target is an AbstractReducedField and the source is either an AbstractDataField, AbstractOperation, or array of some kind. These work on the GPU. In the end the solution was simple since we now subtype AbstractArray; we only need to pass a view into the interior indices of the target to Base.mapreducedim! (which on the GPU ends up at GPUArrays.mapreducedim!).

The result is that AveragedField(op::AbstractOperation) no longer needs to allocate memory for the three-dimensional result of computing op. Instead, op is reduced in a kernel. This is both faster (much much faster, I think --- though a benchmark is a good idea) and more memory efficient. It also greatly simplifies compute!(field::AveragedField):

function compute!(avg::AveragedField, time=nothing)
    compute_at!(avg.operand, time)
    mean!(avg, operand)
    return nothing
end

So, resolves #1422.

I also took the liberty of resolving #1610 and nuking interiorparent...

[tomchor]

Seems like a very positive change

[tomchor]

Don't know if it's relevant, just caught the fact that new_data is being imported from two different modules. Is that correct?

[glwagner]

Not correct -- nice catch!

[tomchor]

Much simpler!