In-place interpolations revisited

[timholy]

By comparison with #34, this is probably a better approach: rather than defining a new boundary condition Interior and then making the indexing rules treat padding differently for this condition, this PR adds a pad parameter to BSplineInterpolation: for instance, pad=0 for anything other than Quadratic, for which it is typically 1 unless (1) using Reflect or (2) using interpolate!. The code generation uses this parameter to compute appropriate index offsets.

The first commit in this PR also improves the type-stability (and adds tests for it) of several key functions. Also, I think the size function was broken (it included the consequences of padding, which I assume is not what you wanted?); this is fixed in the 2nd commit by exploiting the new pad parameter.

[timholy]

With the improvements in filtering, this is nice to see (suitable warmup has been done already):

julia> A = rand(300,300,30);

julia> Ai = copy(A);

julia> @time itp = interpolate!(Ai, BSpline(Quadratic(Flat)), OnCell);
  54.937 milliseconds (198 allocations: 49757 bytes)

julia> Ag = copy(A);

julia> @time yi = InterpGrid(Ag, BCreflect, InterpQuadratic);
 299.404 milliseconds (108 k allocations: 26187 KB, 20.47% gc time)

But this was not so nice:

julia> function interp_all!(dest, itp)
           for k = 1:size(itp,3)
               for j = 1:size(itp,2)
                   for i = 1:size(itp,1)
                       dest[i,j,k] = itp[i,j,k]
                   end
               end
           end
           dest
       end
interp_all! (generic function with 1 method)

julia> B = similar(A);

julia> @time interp_all!(B, itp);
  10.251 seconds      (2700 k allocations: 42188 KB)

julia> @time interp_all!(B, yi);
   1.069 seconds      (4 allocations: 144 bytes)

I'll see if I can fix that.