GLASS: Test-Time Acceleration for LLMs via Global-Local Neural Importance Aggregation