Matrix Compression via Randomized Low Rank and Low Precision Factorization