MV-TER (Multi-View Transformation Equivariant Representations) SELF-SUPERVISED MULTI-VIEW LEARNING VIA AUTO-ENCODING 3D TRANSFORMATIONS