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