enable_thinking

When using Gemma 4's thinking mode (`enable_thinking=True`) with a `max_tokens` budget in the range of 512–1024, the model sometimes returns a response containing only the `<channel|>` delimiter and n