FaSDiff: Balancing Perception and Semantics in Face Compression via Stable Diffusion Priors