[quant] Add reduce_range params for quantized_lstm (#39604)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39604
This change preserves BC for older models that are saved with reduce_range set to false.
Newer models will use the version information in RNN module to toggle reduce_range parameter
Internally this is implemented using a new CellParams type that calls the linear functions with reduce_range option set to true.
New models serialized will use the CellParams struct for the `__getstate__` and `__setstate__` calls. Older models using QuantizedCellParamsDynamic will continue to use their original serialization/de-serialization methods
tested using LSTM BC test and test_quantized_rnn
Test Plan:
python test/test_quantization.py
Imported from OSS
Differential Revision: D21977600
fbshipit-source-id: 0cb0e098b87207b537574d3beeab1f341c41c0d2
