GPU Assembly AtPoints

[jeremylt]

This PR adds AtPoints assembly kernels to */gen

• Cuda Diagonal
• Hip Diagonal

And then for the next PR I'll do

• Cuda Full
• Hip Full

[jeremylt]

Marked as ready for review - this passes the Ratel test suite

cc @zatkins-dev because you were interested in the code generation - happy to step through the content with you in a screen share or otherwise

[jeremylt]

Note - we should also test somewhere with Ratel + this branch + HIP before merging

[jeremylt]

Note to me - I forgot to generalize to multiple active bases. Small fix that's easy

1. array const CeedInt num_active_in[2] = {1, 3}; (or however many, whichever ind)
2. array const CeedInt active_r_e_in[2] = {r_e_in_1, r_e_in_3};
3. loop for (CeedInt b = 0; b < num_active_in; b++) {
4. update n loop as needed

[jeremylt]

Ok, that plan above actually won't work because of templates. I'm restricting this to just when we have a single active basis, which means we'll get best performance if we fieldsplit when we have multiple active bases, which is the plan anyways so we can use pMG on the displacement block.

[zatkins-dev]

In that case, I think this is good!

[jeremylt]

Do you have a machine you can check this branch with Ratel on quickly?