support images in tool call output with OpenAI

[simonpcouch]

This is a revival of #617.

In the docs for the completions API, the link to function calling heads to this page. Under "Tool call outputs - output we generate for the model", I see:

"The tool call output might return a JSON object (e.g., {"temperature": "25", "unit": "C"}, indicating a current temperature of 25 degrees), Image contents, or File contents."

That linked "Image contents" page is just the docs for the usual entry point from that used in user chat input.

I decided to give a go at using the same interface while working on an eval today, and am seeing convincing results; only counted as ~1000 tokens and reasonably strong visual understanding.

[simonpcouch]

Ah, well, my Claude-Coded implementation assumes a tool return value based on the Anthropic spec rather than a ContentImageInline, so probably isn't very useful for yall.

[simonpcouch]

I had forgotten that #735 exists--that implementation work like a charm. Will close in favor of that PR.