Set tokenizer model_max_length to 2048

As described in the FLAN-UL2 blog, the receptive field of the model was increased from 512 to 2048.

There is also a `n_positions` in the model config, set to 512, but I can't see its use in `transformers` 🤔

tokenizer_config.json

@@ -103,7 +103,7 @@
   ],
   "eos_token": "</s>",
   "extra_ids": 100,
-  "model_max_length": 512,
+  "model_max_length": 2048,
   "name_or_path": "google/ul2",
   "pad_token": "<pad>",
   "special_tokens_map_file": null,