Implement "sanity checks" training runs

[AmitMY]

We should implement a few training configuration files to train small models to do:

1. String repetition (--dataset_text_template "<text> {text} <again> {text}")
2. String repetition + (--bytes_encoder_model_name_or_path None)
3. Word deconstructions ("Strawberry S1 T1 R3 A1 W1 B1 E1 Y1") - lots of words, maybe only english

and evaluate actually in generation.

[AmitMY]

Implemented:

1. https://github.com/sign/WeLT/blob/main/training/experiments/easy-tasks/string-repetition.yaml
2. https://github.com/sign/WeLT/blob/main/training/experiments/easy-tasks/ocr.yaml