fix(openai): parsing of streaming completions

[Eisfunke]

streamCompletion previously failed often (not always though), because parseResponseChunk assumed that a chunk of the input stream from the OpenAI api would only ever contain a single message. However, the API often sends two data messages in one chunk separated by two newlines (which the format allows) together at the start of a completion, which failed to parse as JSON. Because the unparsed chunks were stuck in the buffer, they were always appended to the beginning for all other chunks, which therefore also didn't parse, leading to all messages not being parsed.

I fixed this by re-chunking the stream into lines with Streams.lines. And also added parsing logic for comments, multi-line messages and messages separated by double-newlines, as specified in the Server-Sent Events spec the API uses here. It seems the API currently doesn't send comments and only ever sends single-line messages, but in case they change that, IHP can now handle it.

https://platform.openai.com/docs/api-reference/chat/create#chat/create-stream
https://developer.mozilla.org/en-US/docs/Web/API/Server-sent_events/Using_server-sent_events#event_stream_format