Eliminate array copy in NormalizeOp

[am15h]

[xunkai55]

@googlebot rescan

[xunkai55]

Though it's not really "ready to pull", I'd assign this label to trigger an internal code sync, thus I can benchmark the change. Please stay tuned. Thanks!

[xunkai55]

Per benchmarking, the change has negative impact on performance of a followed Quantization Op in a typical image preprocessing pipeline:

Resize - Crop - Rotate - Normalize - Quantize

And performance of "Normalization" Op is not improved.

At the first glance, I suspect there's something with ByteBuffer.

Let's hold this for a while and I'll try spare some time investigating it.