High-dimensional subset recovery in noise: Sparsified measurements without loss of statistical efficiency