feat(qmoe): support 2-bit expert weights in CPU kernel

[Rishi-Dave]

Summary

• Extend the CPU QMoE contrib op to accept expert_weight_bits=2, unpacking four 2-bit values per byte LSB-first to match the convention used by MatMulNBits.
• Generalize the zero-point pack size formula to 8 / num_bits so 2-bit zero points pack four-per-byte, while preserving existing 4-bit and 8-bit behavior.
• Update the schema attribute description to list the valid values (2, 4, 8).

Motivation

Quark-quantized 2-bit Mixture-of-Experts models cannot currently run inference in ORT because the CPU kernel hard-rejects expert_weight_bits values other than 4 or 8. This change closes that gap on the CPU path.

Fixes #28163

Changes

• onnxruntime/contrib_ops/cpu/moe/moe_quantization_cpu.cc
  • Loosen the expert_weight_bits validation to accept 2, 4, or 8.
  • Add a 2-bit unpack branch ((byte >> ((c % 4) * 2)) & 0x3) alongside the existing 4-bit branch in both the row-wise and block-wise asymmetric paths inside DequantizeBlockWithMlas.
  • Generalize zp_pack_size to 8 / num_bits (yields 4 / 2 / 1 for 2 / 4 / 8 bit; behaviorally identical for 4 and 8 bit).
• onnxruntime/core/graph/contrib_ops/contrib_defs.cc
  • Update the expert_weight_bits attribute description to enumerate the valid values.
• onnxruntime/test/python/transformers/test_qmoe_cpu.py
  • Generalize the row-wise and block-wise quant helpers to take a bits parameter (defaults to 4 to preserve existing tests) and support 2-bit packing.
  • Add 2-bit row-wise and block-wise CPU QMoE test cases.

This intentionally leaves the MLAS LUT GEMM fast path for 2-bit out of scope — the dequant + MlasGemm path is sufficient for correctness, and a follow-up can add the LUT fast path once the surface here is settled.

Test Plan

• Existing 4-bit and 8-bit row-wise / block-wise CPU tests in test_qmoe_cpu.py continue to pass (the helpers default to bits=4).
• New 2-bit row-wise and block-wise tests construct small QMoE models, run the CPU kernel with expert_weight_bits=2, and check the output stays within a loose tolerance of the float reference. The tolerance is widened for 2-bit because the format is intrinsically lossy.
• lintrunner is clean on the changed files.

Local CPU build/test execution was constrained by the dev environment (no prebuilt onnxruntime present and no incremental build dir to reuse), so the new tests are exercised via CI on this PR.

[tianleiwu]

The PR has overlap with #28185