On the Hardness of Learning to Stabilize Linear Systems