fragile docker image / installs

[tommybotch]

Current Docker image is fragile, with the following experiences:

• Removing feature-splatting from requirements.txt seems to mess with opencv-python
• Constraint of "numpy<2.0.0" requires scipy to be pinned at scipy==1.11.4
• ultralytics bumps many packages to newer versions
• Also remember there being a reason why we needed to go back down to torchtyping==0.1.5 but can't find the error...

Changing any one of these breaks the Docker image, or results in needing manual installations. Ideally, we can locate all of these pressure points and centralize the installs within the pyproject file