EoRA: Training-free Compensation for Compressed LLM with Eigenspace Low-Rank Approximation