Feature learning from non-Gaussian inputs: the case of Independent Component Analysis in high dimensions