[0.3.0] Abstracting Pytorch example

[Syakyr]

Closes #17.

This is a big merge before freezing the code to 0.3.0 with a number of changes on top of abstracting the Pytorch example and creating a package-agnostic base template to implement models of the user's choosing.

Changes made:

• Experimental RoCM support
• Package-agnostic base template to build relevant problem templates from
• Cosmetic changes for consistency and/or simplicity
• More explicit methods of deployment depending on use case/availability of infrastructure
• Refined checks with regards to the cookiecutter input
• Removed dockerfiles for JupyterLab and VSCode images (moved to kapitan-hull-admin)

[Syakyr]

Problem with the gpu (model training) Dockerfile such that pytorch is forced to run cpu only due to the dependencies. Would need to find check if nvidia image base is needed by installing the gpu dependencies onto the current cpu image (to reduce image size). Concern is that CUDA components might be installed twice in anaconda namespace, thus increasing the image size for the gpu (model training) image.

[Syakyr]

So installing pytorch-cuda as a dependency on ubuntu:20.04 image would allow pytorch to use CUDA, circumventing the need to use the Nvidia images. Checking whether rebuilding the GPU image using the CPU dockerfile as well as modifying the conda yaml file to install gpu-enabled Pytorch packages would result in model training with GPU, or model training with CPU if GPU is not supplied.