[PyTorch FE] Add 2-bit (u2) weight decompression subgraph support

[ljaljushkin]

Details:

• Added u2_compression_stack pattern matcher in utils_quantize.cpp that detects the NNCF
  unpack_uint2 subgraph: aten::stack([bitwise_and(packed, 3), bitwise_and(bitwise_right_shift(packed, 2), 3), bitwise_and(bitwise_right_shift(packed, 4), 3), bitwise_and(bitwise_right_shift(packed, 6), 3)], dim=-1)
  and replaces it with a single element::u2 constant.
• Added U2ConvertReshape transformation pass that folds Reshape on u2 constants
  into the constant itself (analogous to U4ConvertReshape for u4).
• Added MarkCompressedWeightConstants transformation pass that marks Convert nodes
  consuming u2/u4/i4 constants with disable_constant_folding and mark_as_decompression
  to prevent MOC from expanding compressed weight constants.
• Integrated u2 pattern detection in both translate_stack (TorchScript) and
  translate_stack_fx (torch.compile) code paths.
• Added u2 → u8 type promotion in CPU plugin's transformation_pipeline.cpp.
• No need for a standalone aten::bitwise_right_shift op converter — the pattern matcher
  consumes the entire unpack subgraph at the aten::stack level.

This enables export of NNCF models with INT2SymmetricWeightsDecompressor to OpenVINO IR,
following the same approach used for INT4 decompression patterns (PR #27048).

Tickets:

• Depends on NNCF PR: [Torch] INT2 symmetric decompression support nncf#3971 (INT2SymmetricWeightsDecompressor)
• ref: Support for INT4 decompression patterns from NNCF #27048 (INT4 decompression patterns)

AI Assistance:

• AI assistance used: yes
• *manually converted u2/u4 qwen/qwen3-4b and evaluated in openvino

lm_eval
--model openvino
--model_args pretrained=$IR_DIR
--device cpu
--tasks lambada_openai

model	quant_config	group_size	Lambada, acc	Lambada ppl	Lambada OV acc	Lambada OV ppl
Qwen/Qwen3-4B	FQ_LORA (avg 3bit - mix of 2/4 bit)	128/64	0.5604	9.6826	0.5572	9.6666