End-To-End TensoRF connection

[SimonDaKappa]

This PR covers the implementation of TensoRF as the nerf-worker container in docker. The nerf-worker is implemented similarly to colmaps sfm-worker, consuming nerf-in to put files in data/inputs, and producing nerf-out to send files from data/outputs.

Features:
TensoRF is now the nerf-worker container in docker
Communication between all workers and web-server via rabbitMQ
web-server now publishes finished sfm jobs to nerf worker
nerf-worker now consumes sfm jobs, runs the TensoRF pipeline, and produces the generated model/video
web-server now consumes finished nerf jobs, saves to MongoDB, and responds to web-app get requests for nerf video

[rougejaw]

At the moment, any significant training of the model results in the process being halted, as Pika sends a heartbeat every 300 seconds that is being blocked by the training process, killing the process.

[SimonDaKappa]

Pika and GPU fixes for nerf-worker are now in. This is handled in the following way

OLD:
TensoRF/main.py fork main into
| Flask Process
| nerf_worker Pika / TensoRF Training process

NEW:
TensoRF/main.py fork main into
| FLask Process

Use pytorch.multiprocess
TensoRF/main.py from main spawn into
| nerf_worker pika process. On nerf-in consume:
| create thread to run training and rendered, use threadsafe callback to ack and publish vid to
| nerf-out

Additionally, Aidan has integrated logging into sfm-worker, web-server, and nerf-worker, with some minor tweaks I made to work with my TensoRF changes.

Something to note, colmap seems to spit out images not in the order which we would get them by splitting the video into frames, so the TensoRF training path is very unsmooth. Each frame to the next can rotate +/- 180 degrees around the scene center, leading to the final reconstructed video following the same nonsmooth path.

Reasoning:

1. Nerf flask server can only send get requests, not receive get requests if process spawned. Could
   not figure out why.
2. Since the research code is baaaad, it loads tensors into cuda device memory by redeclaring the cuda
   device every single time. CUDA has a known bug where forked processes, ( I assume something to do
   with trying to copy-on-write cuda objects not working ) cannot declare cuda device more than once.
   Pytorch has a wrapper around python.multiprocessing.spawn to spawn a fresh process and handle
   transfer of tensors in that process. This works with the research code and heartbeat fix for pika.

[SimonDaKappa]

Also added a .sh script empty_data.sh to clear out all the worker/webserver input and output data folders while preserving fild structure