Compressor-VLA: Instruction-Guided Visual Token Compression for Efficient Robotic Manipulation