Mismatch in Predicted Labels and ignored_labels=['O'] Not Working

For my token classification task, the predicted labels from the explainer differ from those obtained using the pipeline. Some tokens that should have a label are instead predicted as 'O' by the explainer. Even after setting `ignored_labels=['O']`, these tokens are still included in the visualization (with only the true 'O' tokens being excluded) and continue to be displayed as 'O' in the visual.

pipeline prediction :

```
[{'end': 50,
'entity_group': 'OrderAndDelivery',
'score': 0.91371477,
'start': 4,
'word': 'colis a été marqué comme livré alors que je ne'},
{'end': 62,
'entity_group': 'OrderAndDelivery',
'score': 0.6080048,
'start': 56,
'word': 'jamais'}]
```
Explainer :


Has anyone else experienced this issue or found a solution?
Below is the code:
```
config = AutoConfig.from_pretrained('models/ner_model_camembert_v7')
max_length=120
model = AutoModelForTokenClassification.from_pretrained('models/ner_model_camembert_v7', config=config)
tokenizer = AutoTokenizer.from_pretrained('models/ner_model_camembert_v7', config=config, truncation=True, return_offsets_mapping=True, padding="max_length", max_length=max_length)

model.eval()

ner_explainer = TokenClassificationExplainer(
model,
tokenizer
)

sample_text = "Mon colis a été marqué comme livré alors que je ne l ai jamais reçu"

word_attributions = ner_explainer(sample_text, ignored_labels=['O'])

pipe = pipeline("token-classification", model=model, aggregation_strategy="simple", tokenizer=tokenizer)
output_model = pipe(sample_text)
pprint(output_model)

ner_explainer.visualize()

```

Start with TokenClassificationExplainer and compare its output with the transformers token-classification pipeline shown in the issue. Trace how ignored_labels=['O'] affects attribution and visualization, using the supplied sample text to reproduce the mismatch. Done means predicted labels align with the pipeline and ignored O tokens are excluded from the visualization.

Written by the indexing model from the issue text.