re-implement sequential fine-tuning

[tmills]

This is currently blocked by an exception, but we should try to reimplement it.

so we need to first:

1. fine-tune on task A and save the model
2. re-run train_system where we load the task A model with AutoModel but delete the feature transform and classifier layers
3. re-save that model to a temporary path with the layers deleted
4. create a new CnlpConfig based on the task B command line parameters
5. create a new model that takes the temporary path from 3 and the config from 4 using the CnlpModelForClassification constructor

we need to replace this constructor call with an automodel call: https://github.com/Machine-Learning-for-Medical-Language/cnlp_transformers/blob/main/src/cnlpt/train_system.py#L461

3 is done here: https://github.com/Machine-Learning-for-Medical-Language/cnlp_transformers/blob/main/src/cnlpt/train_system.py#L471

4. can look something like this: https://github.com/Machine-Learning-for-Medical-Language/cnlp_transformers/blob/main/src/cnlpt/train_system.py#L493

5. can look like this: https://github.com/Machine-Learning-for-Medical-Language/cnlp_transformers/blob/main/src/cnlpt/train_system.py#L504, using the newly created config

[tmills]

Adding @spencerthomas1722 since he has a use case for this feature.