Trade-off between conservation of biological variation and batch effect removal in deep generative modeling for single-cell transcriptomics