FAST: Efficient Action Tokenization for Vision-Language-Action Models