Memory-Side Acceleration and Sparse Compression for Quantized Packed Convolutions