Numerical inaccuracies on consumer grade GPUs

[mikovtun]

Running the test suite on device (#91) fails for the H3 test system (GKS GGA) when running on an NVIDIA RTX 3060 because the calculated VXCx and VXCy norms differ from the CPU calculations by about 1e-7. This failure does not occur on the professional line of cards (tested on A100, L40, and A40). The likely cause is a different implementation of the rsqrt call between the RTX and professional cards.

This is the first time different GPUs have caused a non-negligible difference in calculations, which this issue serves to document. While consumer cards can be used (with caution), they are not the intended hardware to be used with GauXC.