Variable-rate hierarchical CPC leads to acoustic unit discovery in speech