Deterministic equivalent and error universality of deep random features learning