Implement squad validation set preprocessing and scoring

[plpxsk]

Currently, any valid and test datasets (in addition to train) come from the squad "training" batch only.

This is because the squad training batch has exactly one answer, while the validation sets have multiple "correct" answers, and this requires slightly different pre-processing

Implement validation set pre-processing so that:

• the actual validation set can be passed for test/validation
• to compute performance metrics on these sets

See example code in HF: https://huggingface.co/learn/nlp-course/chapter7/7?fw=pt#processing-the-training-data