[Quantization] Make Int8 MaxPool and AvgPool handle different input/output fp32 dynamic ranges

[rdzhabarov]

Currently both operations expect Scales and Offsets for inputs/outputs to be the same.
In reality, most common case would be a different quantization parameters for inputs/outputs.
There is a high chance that Max pool for example would increase the lower bound of possible values and hence output needs to have a smaller fp32 dynamic range.

There are several places which needs to be fixed:

• Implementation of mentioned above functions on CPU and Interpreter
• Quantization procedure which should not enforce same scale/offset for inputs/outputs
• Caffe2 loader of quantized operations for pooling ops

See: #2016 (comment)

[rdzhabarov]

apparently avg pool already supports this. For max pool it'd make sense from the efficiency stand point to keep the same params as inputs.