Asymptotics of representation learning in finite Bayesian neural networks