add support for vision-language model endpoints

[mattf]

add support for vision-language models, those that can accept images and text as input and produce text.

these are akin to https://platform.openai.com/docs/guides/vision with three notable differences -

1. images can be passed with img tags in the regular text content
2. images can be passed as NVCF asset ids
3. not all model endpoints support all features, e.g. server-side download of images not available with adept/fuyu-8b, google/deplot, microsoft/kosmos-2, google/paligemma; some models endpoints restrict image size; some models support one and only image; some models do not support gif or webp; kosmos-2 does not support streaming

prototype support for vlm existed in 0.2.2 and before

breaking changes -

• remove client-side image download w/ resizing and base64 encoding

this will be available in v0.3