The interpolants we in `Interpolations.jl` are described by two arrays:
the knots and the coeffs. When `Adapt` is called on these interpolants,
CuArrays are allocated on the GPU. For large data, this is inefficient.

In this commit, I add a system to avoid these allocations. This is
accomplished by add a dictionary to `InterpolationsRegridder`. This
dictionary has keys that identify the size of the knots and coefficients
and values the adapted splines. When `regrid` is called, we check if we
have already allocated some suitable space in this dictionary, if not,
we create a new spline, if we do, we write in place.

This removes GPU allocations in the hot path (ie, the regridder is used
in a time evolution with always the same data and dimensions), while
also keeping the flexibility of reusing the same regridder with any
input data.