Use httr2::req_perform_stream(round = "line")

[romainfrancois]

closes #63

This uses httr2::req_perform_stream_lines() from r-lib/httr2#437 so that the data from open ai is dealt with line by line instead of some number of bytes.

This PR is meant more as a conversation aid for the httr2 PR.

If httr2 actually had httr2::req_perform_stream_lines() or something serving the same purpose: consume stream line by line, then I believe some of the logic around ch_env$stream$raw could be improved, because we would know that we get complete data: {json} lines, and so openai_stream_parse() etc ... could be simplified.

Compared to #64 which has to do tricks with the bytes buffer, this feels much nicer.

So now we can stream the golem poem, cc @ColinFay 🫣

Screen.Recording.2024-02-08.at.11.58.50.mov

[edgararuiz]

Thanks @romainfrancois !