Call `getQSchemeAndQParamMap` later in `quantizeTensors` (#31406)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/31406
Previously we record quantization parameters for a given value when we collect the observer nodes,
but actually the quantization parameter can vary depending on each module instance, to achieve
that, we need to delay the call to later stage and only record the `Value*` that's needed
in `collectObserverNodesAndValueToQuantize` function
Test Plan:
.
Imported from OSS
Differential Revision: D19162369
fbshipit-source-id: e0f97e322d18a281bf15b6c7bbb04c3dfacb512f
