Support pre-training or model resuming when loading a checkpoint

We should support two behaviors when loading a checkpoint:

• --train.resume option (currently available) to continue training.
• --train.pretrain option to only load the weights (see this script)
  • reset the callbacks (EarlyStopping)
  • reset the optimizer
  • reinitialize the last layer in case of a different alphabet

This will cover two uses cases:

• --train.resume to continue training on the same dataset
• --train.pretrain to load pretrained weights when fine-tuning on another dataset