(Feature Request) Assistant prefill for Anthropic models

[maeeeeee]

Per Anthropic docs, Claude models support adding a prefill that allows you to determine the beginning of the model's response. Adding support for this in Mantella would go a huge way in being able to reduce rejections and filtering; in my experience, something as simple as "Certainly, here's my response: " does wonders.

[art-from-the-machine]

Thanks for bringing this up @maeeeeee! It looks like this might be possible to implement across all LLMs so I will aim to include this in the next release