Deep Clustering with Incomplete Noisy Pairwise Annotations: A Geometric Regularization Approach