Currently we refer to the parameter_axis as the first dimension, but the DPP/cvxpylayers papers puts it as the last (third) dimension. We need to adjust to avoid further confusions. This will also lead consistency for the row and col axis.
In terms of implementation, only the NumPy backend requires major changes. The other ones will probably only involve updating some docstrings.