Implementation of Calvo3 (or Paco's method) in local_fast_trainer

This new method is an extension of the Calvo2 method (using random uniformly distributed samples from 5 images to generate training batches).

According to @fjcastellanos, this is what is needed for Calvo3 (Paco's method):
> The main change that I have to make in the code is to deal with the path of the images instead of preloading all the images before. However, as there are dependencies with Rodan, I 'm not sure how to test the code

@fjcastellanos, based on my understanding of the code, the changes you propose should be implemented on `local_fast_trainer.py` file, in a new branch derived from `develop`.

That script is independent of Rodan and can be trained locally, so it will allow you to iterate quickly over your changes.

@kemalkongar has volunteered to help you oversee those changes if you need it.

Preferably, push your changes to your branch frequently, and we (e.g., @kemalkongar or myself) can jump in to peer-review the code or help if needed.

Start with local_fast_trainer.py and compare the existing Calvo2 batch-generation flow with the issue's path-based Calvo3 requirements. Run the local training workflow to understand how image inputs are currently handled, then verify that the new method can train locally without preloading all images and remains compatible with the stated Rodan dependencies.

Written by the indexing model from the issue text.