A16S8 gemm && tensor-wise quantization (#441)

1. To meet the framework requirement: Zero Point type = int8_t, Scale
Type = fp16/bf16
2. numeric_convert<half_t, int8_t> can not work and walk around by
element-wise data conversion
3. tensor-wise quantization 
4. UT half_t x int8_t
5. Move the default copy from inner loop to outer loop

---------

Co-authored-by: Alejandro Acosta <[email protected]>
Co-authored-by: Joe Todd <[email protected]>

Author: 3 people