KSDiff: Keyframe-Augmented Speech-Aware Dual-Path Diffusion for Facial Animation