Adds optimization from proxy application

[dmclark17]

Ported proxy app changes from my end to GauXC. I am marking as a draft because I did not include the downstream changes from the matrix transpose edit so it is not currently correct.

[wavefunction91]

@dmclark17 I've rebased this locally, do you want me to push directly or have you review in a separate branch?

[dmclark17]

I think pushing directly here would work

[wavefunction91]

@dmclark17 Up-to-date. You can toggle your compact collocation kernel implementation by uncommenting
https://github.com/wavefunction91/GauXC/pull/19/files#diff-c52215967dc1bb88bf7adcefa6c71b05a94b16c7af12c87ff3ea889112716b3cR4

and changing the analogous kernel launch. FWIW, this increases the register usage (-gencode sm_70,compute_70 -O3) from 64 -> 88.

Note, I'd like to either

1. Settle on one (optimized) implementation (i.e. remove the alg variants I have toggled)
2. Determine a criteria (hardware or inputs or both) to dispatch to a particular implementation.

[dmclark17]

For the collocation_device_masked_combined_kernel_deriv1 kernel, I am also seeing that the grid-stride version is slower (135ms vs. 152ms for Taxol). I think we can stick with your implementation as the increased register usage is detrimental.

[wavefunction91]

I'm wondering if this is might be due to the fact that we're striding on multiple dimensions, it might be possible that a stride along x (i.e. the grid points) might be beneficial while stride along y (shells) is detrimental, or visa versa.

[wavefunction91]

@dmclark17 Is there anything else you'd like to include in this PR?