FP8 types support in NNCF graph building

[alexsu52]

Changes

• Added support for "f8e4m3" and "f8e5m2" types in NNCF graph building.
• Extended TensorDataType to TensorDataType.f8e5m2, TensorDataType.f8e4m3, TensorDataType.nf4

Reason for changes

Support for compression of models with fp8 and nf4 weights.

Related tickets

ref: 164161

Tests

test_compare_nncf_graph_precision_synthetic_models

[nikita-savelyevv]

Which usage scenario does this PR enable? Compression/quantization of already fp8-quantized models?

[jane-intel]

This PR enabled us to convert DeepSeek R1 model (with original f8 weights) via optimum intel. Thank you.

[nikita-savelyevv]

If we claim FP8 model support for weight compression, we should add corresponding tests. For example here: tests/openvino/native/quantization/test_weights_compression.py::TestActivationWeightDtype.test_compression_for_different_dtypes.

As I understand only fp8 support is required for now. I would suggest not to claim nf4 support just yet because some additional effort is required to enable it. This is because Tensor.reshape() needs to be implemented for nf4 nncf Tensors in ov backend.

FP8 compression should work after alexsu52#20

Edit:
Discussed in alexsu52#20

[alexsu52]

Which usage scenario does this PR enable? Compression/quantization of already fp8-quantized models?

Skip fp8 weights during compression. Before this fix, NNCF throw a runtime error: