Hardware-Efficient Compression of Neural Multi-Unit Activity