[DRAFT] Add smooth quant

[mhs4670go]

This commit adds smooth quant algorithm.

ONE-DCO-1.0-Signed-off-by: seongwoo [email protected]

[mhs4670go]

@jinevening
Overall, let me describe changes briefly.

1. Pattern file saves act_scale instead of scale.

For smooth quant, it saves activation max first and, for now, "scale" for smooth quant is calculated in runtime.

2. fm-equalize can receive each drivers path manually. Of course default path is same with the current one: same directory of fm-equalize.

This is for flexibility and testing (fm-equalize-value-py-test).

3. Type of patterns is only "ScaleOnly".

4. Passes are only "FusePreScalePass" and "FusePostScalePass".