Make coca and HF work together

[gpucce]

@ae86280 the HF support is indeed not good for coca yet this PR is to address this and also fix #445 @fedshyvana this will also remove the two unused args #434 while adding options for HF @vturrisi this is the PR where I plan to add the option to output the raw image tokens in CoCa #458

I should be able to work on it in the next days and fix this. So far there is nothing much.

High level Todos are:

• [ ] add causal masking for roberta
• [ ] add gpt2 as decoder
• [ ] adjust configs
• [ ] add option to output image tokens in coca visual encoder

[jqyin]

is there still a plan to support coca + HF models ?