Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Tree Search