NEGEMMLowpMatrixMultiplyCore: per-channel quantization support

[alvoron]

NEGEMMLowpMatrixMultiplyCore supports per-tensor quantization only.
Feature request is to support per-channel quantization.

Reproducer: eshoguli@49a91f3

Issue is reproduced if per-channel quantization is changed to per-tensor quantization in the provided example:

   // per-tensor quantization:
   const QuantizationInfo src2_qinfo(0.2f);

   // per-channel quantization:
   const auto scales2 = generate_quantization_scales(6, 0.2f);
   const QuantizationInfo src2_qinfo(scales2);
   std::cout << "scales2: " << scales2 << std::endl;