DG: Add early stopping support to LightningModel

[dagarfield]

Description

train_params accepts early_stopping, patience, monitor, and mode keys but they were never wired up. This PR connects them to a pytorch_lightning.callbacks.EarlyStopping callback in train_on_dataset.

Changes in src/grelu/lightning/__init__.py:

• Import EarlyStopping from pytorch_lightning.callbacks
• Add early_stopping, patience, monitor, and mode to default_train_params (disabled by default)
• Conditionally append an EarlyStopping callback in train_on_dataset when early_stopping is True

Related Issue(s)

N/A

How to test

Pass early stopping params in train_params and verify training stops when the monitored metric plateaus:

train_params = {
    "task": "regression",
    "loss": "mse",
    "early_stopping": True,
    "patience": 3,
    "monitor": "val_pearson",
    "mode": "max",
    ...
}
model = grelu.lightning.LightningModel(model_params=model_params, train_params=train_params)
trainer = model.train_on_dataset(train_dataset, val_dataset)
# Training should stop early if val_pearson doesn't improve for 3 epochs

[dagarfield]

On this same front, any thought for about adding lr_scheduler or similar?
Someone may have just had a bad time with a dilated conv model...