serve.openai_api_server.generate_completion_stream fails on long (~4000+ token) responses

Nobody has claimed this yet.

[This code](https://github.com/lm-sys/FastChat/blob/main/fastchat/serve/openai_api_server.py#L579) assumes that each chunk received from the worker will consist of one or more complete dicts.

With default settings, this stops being the case for long responses, leading to a crash as `data = json.loads(chunk.decode())` tries to load a non-dict byte-string.

See [this](https://chat.openai.com/share/ac3c6ec7-0f6f-4fef-938d-04fd755e788f) explanation. Not the part about increasing BUFFER_SIZE, that's just kicking the can down the road, but rather the part about accumulating data into a buffer until a `delimiter` is reached.

Start in fastchat/serve/openai_api_server.py at generate_completion_stream around line 579, and reproduce the failure with a response longer than about 4,000 tokens. Trace how worker chunks are decoded and split, then verify that data is accumulated until the delimiter so split dictionaries no longer crash JSON parsing.

Written by the indexing model from the issue text.