Adaptively adding and removing classes

[ClashLuke]

In response to #24, it'd be a great idea to adaptively handle class creation and deletion.
On a UI perspective, the idea is that an admin can simply add or remove classes (as requested in #24) without caring about the inner details of the model used.
In case of class deletion, this should work without retaining the model. In case of class addition, minor retraining is okay.

Regarding the implementation, this issue is blocked by #2 as PyTorch is a requirement. Using PyTorch the final layer, a conventional linear module, can instead be replaced with a custom linear module.
This custom linear module should contain the features of a classic one, but also needs the "add output" and "remove output" options.