Gravar-mail: Fundamental bounds on learning performance in neural circuits