AtPoints Operator Rebuild API

[jeremylt]

For our AtPoints CeedOperators, we really should have an API to rebuild the operator from new CeedRestrictions. For a CPU backend it will likely just be cheaper to destroy and recreate the operator with new AtPoints restrictions, but for the GPU we could save a fair amount of JiT time this way.

The API could look something like this:

CeedOperatorAtPointsRebuild(op_old, at_points_rstr_new, *op_new)

CeedOperatorCompositeAtPointsRebuild(op_old, *at_points_rstrs_new, *op_new)

[zatkins-dev]

A note from another issue (#1784) that touched on this -- we could rewrite our kernels such that max points per element is a function parameter instead of a compile time constant. It's the outermost loop (and we don't know its length anyway), so unrolling it well is probably not happening right now. That would remove the need to recompile in most cases. Then, we would just need to reset the restrictions with the updated points arrays and indices.

[jeremylt]

I don't think that works quite as you expect if i recall correctly - element size is deeply required to be known at compile time for gen

[zatkins-dev]

Hmm let me take a stab at it and see where the issues emerge.

[jeremylt]

basically, it breaks every restriction at points

[zatkins-dev]

I believe I have it working here #1990 -- it doesn't break as much as you'd expect. I've tested with all of the Ratel MPM tests too.