TARo: Token-level Adaptive Routing for LLM Test-time Alignment