Benchmarks of Existing Operations

Currently we haven't built a benchmark framework to compare the operations in Numba Kernels with RAW CUDA C++. One pending reason why we haven't done so is that we know there will be a performance gap awaiting LTO support. However, this shouldn't be a blocker for us to build the benchmark framework.

There are two aspects of performances that we want to capture by the benchmarks:

1. Infrastructure to benchmark the performance gap between releases, this indicates the performance gain we get from optimizing Numbast, Numba, CUDA over time
2. Infrastructure to benchmark the gap between native CUDA C++ and Numba kernels. This measures the overhead of the additional wrappers we built with Numbast (extra shim, IRs).