A joint deep learning model enables simultaneous batch effect correction, denoising, and clustering in single-cell transcriptomics