What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formation