Add `monitor` as a training hyperparameter to `training_config.yaml`

[jeipollack]

We may want to explore monitoring different metrics when determining which model checkpoint to save during training.

Currently, the code conditionally monitors:

• "mean_squared_error" when using Keras' built-in MSE (typically with unmasked data)
• "loss" when using our custom MaskedMeanSquaredError loss function (which embeds masking logic within y_true)

This logic is currently hard-coded, which limits flexibility when experimenting with alternative loss functions or evaluation metrics.

Recommendation:

Add a new configurable parameter, monitor, to the training_config.yaml file to specify which validation metric to monitor in the ModelCheckpoint callback.

Benefits:

• Enables switching between "val_loss", "val_mean_squared_error", "val_masked_mean_squared_error", etc.
• Avoids the need to modify source code when changing loss/metric behavior
• Aligns with standard Keras patterns by exposing monitor as a top-level training hyperparameter