Add CeedBasisApplyAdd

[jeremylt]

WIP, GPU impl and CeedBasisApplyAddAtPoints to come

[jeremylt]

Ok, design decision point - the usage really only makes sense for CEED_TRANSPOSE especially for why I need this, so I made that an explicit requirement.

I think for GPU I think I'll want separate Apply vs ApplyTranspose kernels in some cases?

[jrwrigh]

the usage really only makes sense for CEED_TRANSPOSE especially for why I need this, so I made that an explicit requirement.

Agreed that having the transpose as an input, but verifying it to always be transpose works well. Might also include that in the docstring for it, ie:

  @param[in]  t_mode    @ref CEED_TRANSPOSE to apply the transpose, mapping from quadrature points to nodes. CEED_NOTRANSPOSE is not valid for CeedBasisApplyAdd.

[jeremylt]

Update - CUDA/HIP ref and shared are good, need to tackle MAGMA

[nbeams]

I'm not able to follow libCEED activity as closely these days, but please ping me if you run into any issues with the MAGMA backend.

[jeremylt]

Thanks! The TLDR here is that I want to sum into the target vector with CeedBasisApplyAdd for CEED_TRANSPOSE for CEED_EVAL_INTERP and CEED_EVAL_GRAD. It looks not too tricky for MAGMA at a glance but I haven't dug in on it

[jeremylt]

Only tensor product grad is correctly summing into the target vector with my changes for MAGMA, unclear why but I'll dig more Monday

[nbeams]

Did a quick review, found some copy-paste bugs

[nbeams]

As a side note, looking at the new tests in this PR reminded me that we might want to check on the test coverage for the non-tensor basis in MAGMA. Now that we have two options -- the RTC kernels for lower orders and standard library calls for the rest -- I'm not sure how often, if at all, the "standard" route is being tested. (For t363 I manually lowered MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_P/MAGMA_NONTENSOR_CUSTOM_KERNEL_MAX_Q to make sure it also passed with the non-custom-kernel path -- and it did -- but of course that won't happen in the usual testing).

[zatkins-dev]

Overall I like this refactor. I noticed a few places where it is probably more clear to index an array instead of doing pointer arithmetic, but if you disagree that's fine.