Support weight quantization

[juncaipeng]

Fix #2809
Related paddle PR

Details:

• add WeightQuantizationPreprocessPass to process scale, expand it to a list
• update conv_bn_fuser to support weight quantization
• transform weight from int8/16 to fp32 in light_api.cc

Base on 500 images, the test results are as follows:

model	quantized op	size	top1	top5	top1 diff (int8/16-fp32)	top5 diff (int8/16-fp32)
mobilenetv1 fp32	-	17M	73.8%	93.8%
mobilenetv1 int8	conv2d mul	4.6M	71.8%	92.6%	-2.0%	-1.2%
mobilenetv1 int16	conv2d mul	8.8 M	73.8%	93.8%	0%	0%
mobilenetv2 fp32	-	14.4M	75.6%	93.0%
mobilenetv2 int8	conv2d mul	4.2M	73.8%	93.0%	-2.2%	0%
mobilenetv2 int16	conv2d mul	7.6M	75.6%	93.0%	0%	0%
resnet50 fp32	-	118.8M	79.4%	95.2%
resnet50 int8	conv2d mul	26.1M	78.8%	94.8%	-0.6%	-0.4%
resnet50 int16	conv2d mul	51.8M	79.4%	95.2%	0%	0%

[yiicy]

LGTM