Tide: Token-Informed Depth Execution for Per-Token Early Exit in LLM Inference