Implement Presto finetuning for croptype task

[cbutsko]

Need to implement the following functionalities:

• [x] finetuning Presto for multiclass target (as opposed to current binary)
• [x] running downstream sklearn models for multiclass target
• [x] making the finetuning classes and downstream classes flexible (for example, we want to finetune using the most granular classes available, but in dowstream task we only need a limited/aggregated classes)
• [x] implement possibility to load the finetuned model and start downstream task from there
• [x] align metrics collection for easier handling of results of multiple experiments
• [x] introduce model naming convention

[cbutsko]

This branch attempts to cover all the described issues https://github.com/WorldCereal/presto-worldcereal/tree/croptype Waiting for review.