On the Hardness of Sparsely Learning Parity with Noise