Use statically sized interpolation stencils and unroll loops over them

[rafaqz]

This line is the majority of run time here:

out .+= product .* view(u, J...)

(it will use runtime iteration in a loop internally)

But we usually know a lot about J at compile time. Instead of taking a view we should extract a StaticArray over static ranges, and move those ranges over the grid using a runtime offset. This can be very fast for small windows as the compiler can heavily optimize the unrolled maps over static arrays/ranges.

See Stencils.jl for an example of this in a similar context. Mostly due to StaticArrays windows, mapstencil outperforms ImageFiltering.jl view based approach by an order of magnitude (ignoring the gains from using KernelAbstractions.jl).

(I can PR this after #33 is in)

[rafaqz]

Practically this means moving degree to a type parameter of BSplineInterpolation.

[rafaqz]

And actually - the end result is in changing the broadcast into out to all happen at once for all stencils rather than adding to a zeroed out array. This should be a very large performence improvement.

Likely it needs some kind of @nexprs or @generated function to unroll the ND iteration, or maybe just a map over a StaticVector of indices (which unrolls).

[rafaqz]

Testing this by manually unrolling it looks like ~2.7x faster for a 3d 101010 array with one NoInterpDim and 2 LinearInterpDim. For indexing a single value its 6.5ns unrolled vs 12ns with the loop so 2x better.

So pretty clearly worth unrolling.