Add optional flag in the text model config to return embeddings for all tokens in sequence instead of just the EOS embedding.

[mkaic]

This is a small QOL PR that implements an optional parameter.

Currently, there is no way to have the CLIP text model return the entirety of its (batch x seq_length x embed_dim)-shaped output tensor, something that's quite useful if you want to use your model as part of, say, a Stable Diffusion training run.

I've added an optional parameter, return_all_embeddings, that can be included in the text_cfg section of a model config. It defaults to False, so this pull request won't change any default behavior. If it's set to True, the model will return the output vectors for all tokens in its input, not just the one for the EOS token.