Feature learning in finite-width Bayesian deep linear networks with multiple outputs and convolutional layers