Profile and optimize hot code paths

We currently have a lot of TODO's in the code around using Numba to accelerate various things like denormalizing actions, g-force calculations, and physics stepping, but I'm not really sure where the slowness is right now. We need to profile, unit test, and optimize hot paths as much as possible to speed up training which already takes a day or so.[ This video](https://www.youtube.com/watch?v=-4tD8kNHdXs) provides a good overview of how to do so with Numba, which is already used extensively throughout the project.

Start by locating the TODOs for denormalizing actions, g-force calculations, and physics stepping, then profile a representative training run to identify the actual hot paths. Review the existing Numba usage and unit-test coverage for those areas. Done means measured bottlenecks are optimized where justified and the affected behavior is covered by unit tests.

Written by the indexing model from the issue text.