Preventing Dimensional Collapse of Incomplete Multi-View Clustering via Direct Contrastive Learning