Change attention_probs_dropout_prob to 0.1 so that triton FlashAttention dependencies are avoided

See [github issue](https://github.com/mosaicml/examples/issues/441#issuecomment-1874817465) for more detail

config.json

@@ -4,7 +4,7 @@
   "architectures": [
     "BertForMaskedLM"
   ],
-  "attention_probs_dropout_prob": 0.0,
+  "attention_probs_dropout_prob": 0.1,
   "auto_map": {
     "AutoConfig": "configuration_bert.BertConfig",
     "AutoModelForMaskedLM": "bert_layers.BertForMaskedLM"