fix(openai_api_server): use exclude_none for completion stream finish chunks

[Chessing234]

Bug

`generate_completion_stream_generator` (`fastchat/serve/openai_api_server.py`) emits the final SSE chunks for a completion stream with `model_dump_json(exclude_unset=True)`, leaving `None` fields (e.g. `logprobs`) in the payload that the preceding comment explicitly says should be stripped.

Root cause

The function ends with:

```python

There is not "content" field in the last delta message, so exclude_none to exclude field "content".

for finish_chunk in finish_stream_events:
yield f"data: {finish_chunk.model_dump_json(exclude_unset=True)}\n\n"
```

The comment and `exclude_*` keyword disagree. The parallel chat-completion path `chat_completion_stream_generator` carries the identical comment immediately above and correctly uses `exclude_none=True`:

```python

There is not "content" field in the last delta message, so exclude_none to exclude field "content".

for finish_chunk in finish_stream_events:
yield f"data: {finish_chunk.model_dump_json(exclude_none=True)}\n\n"
```

This is a copy-paste asymmetry: the completion-stream version kept `exclude_unset` while the chat-stream version was corrected to `exclude_none`.

Fix

Switch the completion-stream finish loop to `exclude_none=True` so it matches its comment and its chat-stream counterpart. One-line change; per-chunk emission inside the loop body (which legitimately uses `exclude_unset`) is untouched.