Fix a typo. The FSDP wrapper should wrap the `MllamaCrossAttentionDecoderLayer`, which was missing.

**Problem**
Currently, we're val_loss.append(best_val_loss) in each epoch. This is misleading because we're appending the corresponding epoch (not best across epochs) quantities in train_loss, train_prep, and val_prep. This is also inconvenient, as one often would like to plot both train and validation losses as a function of the epochs to look for overfitting.

**Solution**
val_loss.append(eval_epoch_loss)

Set numpy seed in finetuning.py to fix it during finetuning (including in custom_dataset.py) and have it set in functions such as Dataset.train_test_split. This avoids having different train/test splits in different ranks, which may cause NCCL collective operation timeout errors.

Co-authored-by: Matthias Reso <13337103+mreso@users.noreply.github.com>

availble -> available