[OV] Add support for nf4_f8e4m3 quantization mode

[nikita-savelyevv]

What does this PR do?

Changes

• Added OVMixedQuantizationConfig for mixed precision quantization scenario. It is initialized with an instance of OVWeightQuantizationConfig and an instance of OVQuantizationConfig.
• Added nf4_f8e4m3, int4_f8e4m3, nf4_f8e5m2, int4_f8e5m2 as possible values of --quant-mode CLI argument. This performs mixed precision quantization, compressing weights to nf4/int4 precision and activations to f8e4m3/f8e5m2.
• Quantization configs refactoring. OVQuantizationConfigBase now contains only model-related parameters. Added to_nncf_dict() method to quantization configs for convenience.
• Renamed OVWeightQuantizationConfig.weight_format to OVWeightQuantizationConfig.dtype and OVQuantizationConfig.activation_format to OVQuantizationConfig.dtype. The latter is done because when OVQuantizationConfig is used, not only activations but also weights are quantized, so activation_format does not correctly represent what actually happens. OVWeightQuantizationConfig.weight_format is renamed for consistency.
• OVBaseModel._prepare_quantization_config() can now create instances of configs other than OVWeightQuantizationConfig.

Examples

CLI

optimum-cli export openvino -m meta-llama/Llama-3.1-8B --quant-mode nf4_f8e4m3 --dataset wikitext2 ./llama-3.1-8b_nf4_f8e4m3

Python API:

model = OVModelForCausalLM.from_pretrained(
    model_id="meta-llama/Llama-3.1-8B",
    quantization_config=OVMixedQuantizationConfig(
        weight_quantization_config=OVWeightQuantizationConfig(bits=4, dtype="nf4"),
        full_quantization_config=OVQuantizationConfig(dtype="f8e4m3"),
        dataset="wikitext2",
    )
)

Some of these changes were implemented thanks to @nikita-malininn .

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]

@nikita-malininn, please take a look as well.

[AlexKoff88]

@eaidova, we hope that this change will not have a negative impact on the OpenVINO Notebooks as it is not backward compatible.

[echarlaix]

yes could make sense to add a warning (+ potentially keep compatibility for one or two releases by setting dtype in case weight_format is provided)

[AlexKoff88]

@l-bat, can you please review this PR as well?

[AlexKoff88]

Overall, it looks good, thanks.

[AlexKoff88]

@IlyasMoutawwakil, @echarlaix, the PR is ready for your review.

[echarlaix]

LGTM, thanks for the addition @nikita-savelyevv

[echarlaix]

yes could make sense to add a warning (+ potentially keep compatibility for one or two releases by setting dtype in case weight_format is provided)

[echarlaix]

Failing tests unrelated so merging, thanks @nikita-savelyevv