[Torch] INT2 symmetric decompression support

[ljaljushkin]

Changes

• Added INT2SymmetricWeightsDecompressor in nncf/torch/quantization/layers.py:
  stores signed int2 values [-2, 1] as unsigned uint2 [0, 3] using zero point of 2;
  four uint2 values are packed per uint8 byte via pack_uint2/unpack_uint2.
• Added pack_uint2 and unpack_uint2 functions in quantize_functions.py for
  packing/unpacking 2-bit values in uint8 containers.
• Updated strip.py to handle INT2 decompressor during model stripping with
  StripFormat.DQ.

Reason for changes

Support INT2 (2-bit) symmetric weight compression of PyTorch models in NNCF,
enabling mixed-precision quantization (e.g., average 3-bit with some layers at 2-bit
and others at 4-bit) for LLM compression with distillation-based QAT+LoRA.

Related tickets

• Depends on OpenVINO PR: [PyTorch FE] Add 2-bit (u2) weight decompression subgraph support openvino#34542
  ([PyTorch FE] u2 decompression subgraph support)
• Ref: [Torch] INT4 weight compression #3014 (INT4 weight compression)

Tests

• tests/torch/quantization/test_symmetric_2bit.py — 15 tests covering
  QuantizeSymmetric and QuantizeSymmetricTorch at 2-bit: scale conventions,
  level count verification, zero-maps-to-zero, gradient flow, input range formulas
• tests/torch/function_hook/quantization/test_weights_compression.py::test_pack_uint2 —
  uint2 pack/unpack round-trip
• tests/torch/function_hook/quantization/test_weights_compression.py::test_int2_symmetric_weights_decompressor —
  INT2 decompressor correctness