Quantised Transforming Auto-Encoders: Achieving Equivariance to Arbitrary Transformations in Deep Networks