T5 conversion script and presets

[mattdangerw]

Draft of weight conversion script for T5 1.1 weights.

Still need to figure out...

• A better approach for approx gelu.
• How to store the weights for the language model output layer.

[dathudeptrai]

@mattdangerw any update ?

[mattdangerw]

@dathudeptrai thanks for the ping. Should be updates on this shortly!

Basically, we were focused on a 0.5 release (just out the a few days ago), with generation utils and simple decoder models.

But that's done, so now we will be full speed ahead on landing T5 and BART, to seed a seq2seq offering.

[abuelnasr0]

Can I contribute in shipping this model?
There is changes I want to make to the T5Tokenizer. there is no changes in this PR for the tokenizer so there will be no conflicts.

1. the tokenizer only checks for the pad_token exists. I will make it check pad_token, end_token, unk_token also.
2. this is written in the code:

# T5 uses the same start token as end token, i.e., "<\s>".

but actually T5 uses pad_token as start token for decoder input. documatation from huggingFace read the decoder_input_ids argument.
3. add extra_ids argument that will add extra ids at the end of the vocabularies for use as sentinels for T5 training.

[mattdangerw]

This is coming in finally on #1277