fix wrong quantization target in weight quantization

[chenbohua3]

This pr contains two things:

1. When bn fold is enable, the folded weight is assign to module.weight, QAT quantizer should quantize module.weight instead of module.old_weight
2. pass the weight that should be quantized to each quantizer directly to make the training graph more clearly.

[QuanluZhang]

could you explain more about this change? weight can be obtained from wrapper, why pass it again?

[chenbohua3]

It's about the code readability. Since we already pass the weight to be quantized (new_weight) to quant_grad here, it is better to directly use it instead of obtaining it from wrapper. The developer can easily know that quantize_weight is a big op to simulate quantization and the op takes origin/bn-folded weight as input. I think using wrapper.weight will make it difficult to understand to structure of the training graph.

[linbinskn]

LGTM. Since this PR also changes quantize_weight in DoReFa and BNN quantizer, please also test them together.

[chenbohua3]

Have added a ut about the interface of quantize_weight