Create a "predict.py" that generally works for TCN

[cameron-a-johnson]

Lesson learned from the 9/22/23 demo:

The latest experiments were testable via PyTorch Lightning's validation pipeline during training, so we've been able to collect performance metrics via training runs.

Then there was friction when pulling out a "predict" function to plug into the live system.

This wasn't done via Lightning for the 9/22/23 demo, because Lightning's predict call was not working properly. So a predict call was instead hand-crafted, and we hit bumps including not normalizing pixel-wise distances property for the TCN's feature vector, using [tlbr] instead of [xywh] to capture bounding box position & size, etc...

But calling any of our trained TCN models via lightning should be possible with one call.
https://stackoverflow.com/questions/65807601/output-prediction-of-pytorch-lightning-model

Let's get that "predict.py" call ready for the next time we need it, ideally in a way that's generalizable to any of our feature vector versions.