quant: Quantization routines only support CPU processing

[cimes-isi]

Quantization routines use unsigned 32 bit integers to pack quantized data. Since PyTorch doesn't support this datatype, we're use numpy arrays before casting back to a Tensor, which currently limits us to running on the CPU. This restriction means we are forced to move data to/from the CPU when sending/receiving data between stages, rather than a more efficient direct network transfer to/from GPUs and NICs.