Refactor ImagingNode to a Normal Class

[MinhxNguyen7]

Summary

• Convert ImagingNode to be a normal Python class.
• The class that owns its own process in a ProcessPoolExecutor, which it uses to do inference.
  • That way, its processing will never block anything. This solves #225.
  • The pool has only one worker, and the class is a singleton to make sure that only one process uses the GPU.
• Log predictions in a thread pool so that it doesn't block.
  • This is probably not necessary since it would only have added a little bit of latency to the inference, and we call get_image_down asynchronously anyway.
• Minorly refactor the now-Perception class and make typing more accurate to get rid of type errors in the class.
• Eliminate CommanderNode busy waiting on perception results.
  • This probably amounted to occupying a whole CPU core unnecessarily for the whole time the Perception pipeline is running.

Test Plan

• Test on the drone.
  • Works. Getting poses, but still with some stuttering.

Issues

• Closes #225

[MinhxNguyen7]

@Dat-Bois This is why the PoseProvider exists. Abstracting the subscription data collection meant that, with dependency injection, I can move the functionality of ImagingNode to another class, and this new class doesn't even need to know that ROS exists. In the future, we can switch out MAVROS for MAVSDK in the implementation of PoseProvider, and the perception module wouldn't have to change at all.