Add performance testing to llm_relay #75
Description
On top of testing for correctness (as we do here . We should also add a test for performance. This ends up being important (and likely expanded on) as once we get a correct kernel it is very important to make the kernel faster as well. Currently, correctness should be saturated (as all pytorch kernels are in an llms training data), however, performance is not, and it is really what we care about.