fix load_in_int8 behaviour for large models

[eaidova]

What does this PR do?

update load_in_int8 logic in from_pretrained for getting the following behavior:

if user specified explicitly load_in_int8=True - model weights will be compressed to int8 after conversion
if user specified explicitly load_in_int8=Flase - model weights will not be compressed
if user does not specify this argument, allow to decide based on model size

This behavior is also aligned with optimum-cli export for openvino

Also changed providing load_in_8bit argument in _from_pretrained method after _from_transformers for avoiding double weights compression during model loading

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[AlexKoff88]

thanks for fixing that