Extend support for the zero-shot text classification task to support encoder-decoder models

Encoder-decoder models can _also_ be used for pure text generation (e.g. Google's FLAN-T5 series of models), and we should support them in the zero-shot text classification task.

Right now we filter supported models via the 'text-generation' tag on the Hub; we'd need to support the 'text2text generation' tag and in these cases load the model with `AutoModelForSeq2SeqLM` instead of `AutoModelForCausalLM`. From the user perspective this should be invisible — the difference between seq2seq and text generation is blurred since encoder-decoder models do both. It means that the zero-shot task will map to multiple HF tasks.