Add functionality to select heads to evaluate error tables on and perform dry runs.

[IsaacParker30]

Added 2 new arguments:

--eval_heads
This allows the user to specify which heads they want to evaluate / print an error table for at the end of training. If not set, will default evaluate all heads.
Usage: Specify all heads you want to evaluate on as a string with heads separated by commas. For example:
--eval_heads='default,pt_head,head3'
Example use scenario would be for replay fine-tuning if the user doesn't want to evaluate the large foundation model database on the pt_head.

--dry_run
Adding this argument will stop the run_train.py script just before the model training is about to begin (calls the tools.train() function. This allows the user to check if they've set their parameters correctly before beginning an expensive training run.

I have tested:

• If dry run works.
• If I can select heads to turn off.
• If code evaluates on all heads if none specified.
• If code stops if user specifies a head to evaluate on that doesn't exist.

Let me know if any changes needed.

[vue1999]

Please fix the pre-commit check suggestions, otherwise looks good.

[alinelena]

@IsaacParker30 this addresses some of the things in #839 should have checked the PR before writing it. I wonder if makes sense to implement also first part in this.

[alinelena]

one small point, if you do not need the head stats at the end you may not want to print it during the training either.

[alinelena]

@IsaacParker30 this will need a small rebase