Feature-based Low-Rank Compression of Large Language Models via Bayesian Optimization