142: Load best model instead of last one

[Johansmm]

Proposed saving/reading of training checkpoints. Steps for correct use:

1. Launch the training. New flags were implemented to complement the saving of checkpoints:
   a. save_top_k : The user can now choose the amount of best_checkpoints to store. He can also decide to save all checkpoints, using --save_top_k -1.
   b. monitor: Metric stored in the name of the checkpoints. Default val_loss.
   c. New storage format : "{epoch}-{step}-{" + monitor + ":.4f}".

2. Perform the checkpoint reading specified from the new flag checkpoint (only defined in mode=eval). Three ways are taken into account:
   a. --checkpoint last (default): Reading of the last checkpoint.
   b. --checkpoint *.ckpt : Reading of a specific checkpoint.
   c. --checkpoint best : Try to read the checkpoint with the best monitor, in two steps:

   1. Direct reading on the file name of the checkpoints in the folder.
   2. Reading all checkpoints (torch.load(ckpt)), extracting the monitor value from : ckpt['callbacks'][ModelCheckpoint]['current_score'].

   If none of the steps is possible, it throws an error.

Note:
With the implementation of 1.c, step 2.c.ii would not be necessary. It was implemented as a transition step between previous projects to this new feature (compatibility).

Closes #142