[quant][graphmode] Insert Observers for dynamic LSTM (#35894)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35894
Insert new TensorListObserver only for weight input of dynamic LSTM
This is because we are currently not observing the activation inputs in graph mode.
Activation tensors are dynamically quantized within the aten::qlinear_dynamic op
Test Plan:
python test/quantization/test_quantize_script.py
Imported from OSS
Differential Revision: D20830387
fbshipit-source-id: 81bd197ee509df41bd7622ed09fa3f199a37573b