Use correct `gelu` function

related discussion: https://huggingface.co/google/flan-t5-xxl/discussions/11

The previous config file was using `gelu` function instead of `gated-gelu` that is automatically set when forcing `is_gated_act`to `True`, more specifically [here](https://github.com/huggingface/transformers/blob/a00b7e85ea4b3e3185440f1f82a6b58e3660b01d/src/transformers/models/t5/configuration_t5.py#L132)
This is not a breaking change since it fixes only for inference. Users that trained a model with `gelu` instead of `gated-gelu` should not be affected by this change. Note that using `gated-gelu` instead of `gelu` can give slightly different results

config.json

@@ -6,13 +6,11 @@
   "d_kv": 64,
   "d_model": 1024,
   "decoder_start_token_id": 0,
-  "dense_act_fn": "gelu",
   "dropout_rate": 0.1,
   "eos_token_id": 1,
-  "feed_forward_proj": "gelu",
+  "feed_forward_proj": "gated-gelu",
   "initializer_factor": 1.0,
   "is_encoder_decoder": true,
-  "is_gated_act": true,
   "layer_norm_epsilon": 1e-06,
   "model_type": "t5",
   "n_positions": 512,