Support streaming responses for accurate token counting

When using streaming (`stream=True`), token counts are not available until the stream completes. Currently `guard.record()` requires known token counts upfront.

Need a context manager or callback that accumulates tokens from stream chunks:
```python
with guard.stream_tracker(model="openai/gpt-4o") as tracker:
    for chunk in stream:
        tracker.add_chunk(chunk)
    # auto-records on exit
```