[multilingual scaling laws] test prediction on hold out prelude checkpoints

Goal
Test our current predictions on prelude checkpoints

Description
We don't have annealed checkpoints of prelude at the moment. To test our predictions (#302), we can fit our N, D parameteric function on stable checkpoints of the multilingual law ladder (#161) and predict for different checkpoints of prelude.
Note that, the validation split for evaluating our multilingual scaling law ladder might be part of the training data of prelude, hence there might be differences in the data splits used for prelude and our multilingual scaling law ladder. However, we assume that training and validation loss are the same and hence will just evaluate all checkpoints (prelude + ladder) on the same validation split used in #161.

Deliverable scope
Predictions vs ground truth validation loss for all prelude checkpoints.

Dependencies
Completion of #161 #308