Ensure --model has to be set when not training from scratch

[gdewael]

I'm adding this pull request directly to the main branch as it does not add extra features.
This PR adds an extra compatibility check between two casanovo options.

Currently, casanovo train with a config.yaml specifying train_from_scratch: False, can still be run without setting the --model parameter. (It then still trains from scratch). I've added a raise ValueError that makes setting --model required when not training from scratch.

[bittremieux]

Thanks for the contribution. Incidentally, train_from_scratch has actually recently been removed, as it was indeed a superfluous config option. See #275. This will be provided in the next Casanovo release.