CLAP: Contrastive Latent Action Pretraining for Learning Vision-Language-Action Models from Human Videos