Apply weight compression after model save to reduce peak RAM during export

[nikita-savelyevv]

What does this PR do?

Currently during OV model export with compression, compression is applied right after the model is converted. Because of this, we can't deallocated the memory required for the conversion before the compression start.

This PR introduces an intermediate full precision model saving step. Model compression is applied after this step allowing to save some memory. For INT8 compression the improvement is estimated to be at least 20%. Please see the figures below for "before" and "after" for INT8 compression of llama-2-7b-hf model.

Memory figures were obtained with https://github.com/openvinotoolkit/nncf/blob/develop/tools/memory_monitor.py

Memory improvement is only for data-free compression via optimum-cli. For export via Python API, i.e. from_pretrained(..., export=True) the memory is already handled efficiently with intermediate saving. The same for data-aware compression via optimum-cli.

Before	After

Model	Peak Mem. Before (MiB)	Peak Mem. After (MiB)	Time Before (sec.)	Time After (sec.)
opt-350m, INT8	1789	1461 (-19%)	20	20 (+0%)
opt-350m, INT4	1734	1466 (-15%)	28	29 (+3%)
tiny-llama-1.1b, INT8	4370	3888 (-11%)	33	32 (-3%)
tiny-llama-1.1b, INT4	4005	3798 (-5%)	63	62 (-2%)
open-llama-3b, INT8	11203	7822 (-30%)	48	55 (+15%)
open-llama-3b, INT4	9750	7822 (-20%)	150	158 (+5%)
llama2-7b-hf, INT8	21108	14551 (-31%)	80	94 (+18%)
llama2-7b-hf, INT4	18039	14531 (-19%)	276	292 (+6%)

For some unknown reason the results for tiny-llama-1.1b are unexpected. There is no yet an explanation for that. For other models there is noticeable memory saving with the cost of longer conversion due to intermediate model saving.

Related tickets
147935

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?

[nikita-savelyevv]

@eaidova @AlexKoff88 please take a look

[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.

[eaidova]

@IlyasMoutawwakil @echarlaix could you please merge?