LLaVA-Octopus: Unlocking Instruction-Driven Adaptive Projector Fusion for Video Understanding