Computational-Statistical Gaps for Improper Learning in Sparse Linear Regression