Statistical inverse learning problems with random observations