Adds warning message when deterministic training loss stagnates too quickly in partial BNNs

[sarah-allec]

Context

Certain deterministic NN hyperparameters may cause overfitting that manifests in the training loss initially decreasing very rapidly and stagnating early. A warning message to the user with suggested solutions would be helpful. Closes #11

Description

After the first epoch, the change in training loss is monitored and any time it drops more than 25% (see figure for justification of choice of 25%), a warning message is printed to the user:

UserWarning: The deterministic training loss is decreasing rapidly - learning and accuracy may be improved by increasing the batch size, adjusting MAP sigma, or modifying the learning rate.

Changes in the codebase

1. Added a function called monitor_dnn_loss in neurobayes/utils/utils.py that prints warning when loss has decreased by 25% at any epoch.
2. Added a call to monitor_dnn_loss in flax_nets/deterministic_nn.py in the training loop (DeterministicNN.train()).

[ziatdinovmax]

Looks good overall. A few suggestions:

• Please clarify the monitor_dnn_loss function's return behavior - either add an explicit return value or document that it intentionally returns None.
• Please Include a length check (if len(loss) > 2:) to avoid potential IndexError with np.diff for the edge case when loss has fewer than two elements.

[sarah-allec]

I implemented the suggestions - thank you! If everything looks good, I will submit the PR.

[ziatdinovmax]

Looks good! There is a seemingly unrelated issue with python-3.10 tests failing, which I will need to figure out later, but this one is ready to be merged.