Training Reproducibility

Thank you for putting together this awesome model!

I'm currently evaluating it in various benchmarks against other models (e.g., deberta-v3) using `AutoModelForSequenceClassification` for fine-tuning, and I'm having an issue with model reproducibility. Specifically, using existing code that allows for finding optimal hyperparameters and then subsequently refitting the final model, I am unable to make my ModernBERT model reproducible using coding that works fine for other model types (e.g., deberta-v3, roberta).

The process of refitting the model uses the huggingface recommended [`model_init`](https://discuss.huggingface.co/t/fixing-the-random-seed-in-the-trainer-does-not-produce-the-same-results-across-runs/3442/2) approach for reproducibility:

```python
final_trainer = Trainer(
args = final_args,
data_collator = data_collator,
model_init = model_init,
train_dataset = dataset_dict_tokenized['train'],
eval_dataset = dataset_dict_tokenized['val'],
compute_metrics = compute_metrics
)
```

However, while this works for all other models, it doesn't seem to for ModernBERT. Initially I thought this might be a flash-attn related issue, but it happens regardless of whether I use that or not.

Any advice here is much appreciated!