[TorchFX] Use torchao for quantize_pt2e API when possible

[daniil-lyakhov]

Context

torch.ao directory is being moved to a separate repo, torchao, and the legacy torch.ao implementation was deprecated in the latest release of PyTorch (see details here)

The solution in our side is to

• Deprecate OpenVINOQuantizer in nncf leaving only the ExecuTorch implementation
• Eventually remove all dependencies on torch.ao from the nncf.quantuze TorchFX backend
• Introduce torchao dependency for the quantize_pt2e API or remove all dependencies on torch.ao from the quantize_pt2e, torch_ao_adapter as well

This PR does not achieve the goal, but makes necessary first steps to achieve the goal

Changes

• OpenVINOQuantizer, TorchAOQuantizerAdapter and quantize_pt2e are using torchao classes whenever it possible (using the conditional import)
• torch_fx_MinMaxBackend and TorchFX transformations don't use the torch.ao FakeQuantize class anymore. Instead, a structure TorchQDQParameters is introduced in src/nncf/experimental/torch/fx/quantization/qdq_parameters.py
• TorchFX transformations.py dependency on torch.ao is resolved (by moving _fuse_conv_bn_ import to other files and moving create_getattr_from_value function to the nncf transformations.py file)
• XNNPACKQuantizer is removed from the tests as the actual torchao implementation is moved to ExecuTorch

Reason for changes

• To support OpenVINOQuantizer from ExecuTorch in quantize_pt2e
• To eliminate dependencies to torch.ao from the transformations.py

Related tickets

170072

Tests

test_openvino_quantizer_with_torch_ao_convert_pt2e is enable only for the torchao implementation

[AlexanderDokuchaev]

Please reuse

nncf/src/nncf/torch/utils.py

Line 416 in cc935e4

def get_model_device(model: torch.nn.Module) -> torch.device:

[AlexanderDokuchaev]

Suggested change

	:param quant_min: Minimum quant value.
	:type quant_min: int
	:param quant_max: Maximum quant value.
	:type quant_max: int
	:param scale: Defines the scale factor used for quantization.
	:type scale: torch.Tensor
	:param zero_point: Specifies the quantized value to which 0 in floating point maps to.
	:type zero_point: torch.Tensor
	:param is_per_channel: Whether quantization is applied per channel.
	:type is_per_channel: bool
	:param ch_axis: Channel axis used for per-channel quantization.
	:type ch_axis: int
	:param quant_min: Minimum quant value.
	:param quant_max: Maximum quant value.
	:param scale: Defines the scale factor used for quantization.
	:param zero_point: Specifies the quantized value to which 0 in floating point maps to.
	:param is_per_channel: Whether quantization is applied per channel.
	:param ch_axis: Channel axis used for per-channel quantization.

:type in docstring used only for API objects

[AlexanderDokuchaev]

Not found where value is not a torch.Tensor, is it really need to use Any?

[AlexanderDokuchaev]

https://github.com/openvinotoolkit/nncf/blob/develop/docs/styleguide/PyGuide.md#33-nestedlocalinner-classes-and-functions