Equivariant Unsupervised Object Detection with Learnable Riesz Transform and Composite Spatial Transformers