feat: use RoBERTa tokenizer to (hopefully) fix some tokenization problems for token classification tasks

tokenizer_config.json

@@ -935,7 +935,7 @@
   "model_max_length": 8192,
   "pad_token": "[PAD]",
   "sep_token": "[SEP]",
-  "tokenizer_class": "PreTrainedTokenizerFast",
+  "tokenizer_class": "RobertaTokenizerFast",
   "model_input_names": [
     "input_ids",
     "attention_mask"