Frequency-Aware Token Reduction for Efficient Vision Transformer