Clearer validation errors

[pfebrer]

With the last documentation refactor, documentation became better and more streamlined, but some validation errors became terrible to read.

The main issue is that when Unions fail to validate, there is no well-defined way of consistently telling which option was closest to passing, and therefore pydantic by default prints the reason why all the possible options failed. This affected mostly error messages on the dataset and target types, which are the most convoluted inputs in metatrain.

This PR:

• Adds discriminators to help pydantic understand which error to raise, reducing the amount of validation errors.
• For architecture options validation, it separates model from training errors so that it is visually easier to distinguish the errors.
• Whenever possible, adds links to metatrain documentation that might help the user.