Allocate the output array out of the loop. This way we can do it once
only when evaluating the spherical harmonics. Pre-compute the square
roots of integers that we use in the loops. Doesn't work well if the
square roots are calculated at the module level (probably a numba
thing).

Results are bad beyond that even with the scaling

Using the BGS API to make some grids of the IGRF to test against their
results.