Update on "Adding tests for save/load support"

Summary: we are able to save a model quantized with a tensor subclass,
save the state dict, then later, load model as meta tensor (i.e. only
load tensor metadata not actually parameters) apply quantization api,
and then load the quantized model state dict.

We change the dtype of the subclass to match the dtype of the
dequantized form, both to align with subclass design guidelines and to
make this work

Test Plan: python test/test.py

Reviewers:

Subscribers:

Tasks:

Tags:

[ghstack-poisoned]

Author: HDCharles