onnxruntime

37b863c6 - Extend DQ→MatMulNBits fusion to support Gemm + per-tensor/per-channel quantization (#27769)

Extend DQ→MatMulNBits fusion to support Gemm + per-tensor/per-channel quantization (#27769) Extends the QDQ `DQMatMulToMatMulNBits` fusion to handle additional quantization patterns beyond the existing blockwise DQ→MatMul case. ### New support - **Gemm**: Fuses DQ→Gemm (with optional bias, including DQ bias) into MatMulNBits, stripping Gemm-specific attributes (`alpha`, `beta`, `transB`). - **Per-tensor & per-channel quantization**: Expands scalar/1D scales and zero-points into block-quantized format expected by MatMulNBits. Block size is configurable via `session.qdq_matmulnbits_block_size` (default: 32). ### Changes - **Selectors** (qdq_selectors.cc): Replaced `ValidateBlockwiseDQForMatMulNBits` with `ValidateDQForMatMulNBits` supporting all three quantization modes. Added Gemm-specific validation. - **Actions** (qdq_actions.cc): Added scale/zp expansion for non-blockwise cases, Gemm attribute cleanup, and bias wiring to MatMulNBits input 5. - **Registration** (qdq_selector_action_transformer.cc): Registered `Gemm` alongside `MatMul`; threaded `qdq_matmulnbits_block_size` from session config. - **Tests** (qdq_matmulnbits_transformer_test.cc): Added tests for per-tensor, per-channel, Gemm (no bias, constant bias, DQ bias), block size options, and negative cases.