Reduce number of specialized methods in stencil kernels

[charleskawczynski]

This is another peel off from #2040.

[charleskawczynski]

Interesting, this seems to be needed for inference. Ah, there was a double space assignment.

[charleskawczynski]

Interesting, this still fails in some relatively complex cases (e.g., our wfact opt examples). I think that this revealed that some of the functions in our window_bounds seems to either require inlining, or function barriers to pass JET tests, which reminds me of #1871.

[charleskawczynski]

One thing I definitely found is that we need this definition:

strip_space_args(args::Tuple{Any}, space) =
    (strip_space(args[1], space), )

[charleskawczynski]

Perhaps map((arg, width) -> arg - width[1], args_idx, widths) needs to be unrolled?