FlexGlue: disable MatMulAddFusion on CUDA (faster Linear layers) - #4914

Stacked on #4908 (FlexGlue matcher) and #4907.

Summary

ONNX Runtime's MatMulAddFusion turns each Linear layer of the FlexGlue matcher (MatMul + bias Add) into a Gemm. Its CUDA kernel first broadcasts the bias into the output with an extra rank-1 GEMM (k = 1), which cuBLAS runs on a slow non-tensor-core kernel (ampere_fp16_sgemm_fp16_128x128_nn, 162 launches per pair, 2.2 ms of GPU time), and only then runs the real tensor-core GEMM with beta = 1.

This PR disables the fusion for the FlexGlue matcher on CUDA. The Linear layers then run as a tensor-core MatMul (beta = 0) plus one fused broadcast Add.

• ONNXModel gets a disabled_optimizers argument, which sets ONNX Runtime's optimization.disable_specified_optimizers session config entry.
• FlexGlue passes {"MatMulAddFusion"} when the matcher runs on CUDA. Nothing changes on CPU or CoreML.

Can the bias be folded into the GEMM epilogue instead?

Not from the model side:

• ONNX Runtime has no operator that fuses a plain bias into the GEMM epilogue. FusedMatMul has no bias input, and GemmFastGelu fuses only the tanh-approximated GELU.
• The bias can't be folded into the following ops either: they are LayerNorm, the attention (after the q/k LayerNorm), and residual adds.

The separate broadcast Add that remains costs 1.36 ms per pair.

Results

Kernel profile: fp16 matcher, N = 2048, A100, ONNX Runtime 1.23 (CUPTI). GPU time: 30.42 → 28.91 ms per pair.

• The 162 SIMT GEMMs (2.19 ms) are replaced by 162 broadcast Add kernels (1.36 ms).
• The tensor-core GEMMs get faster with beta = 0: 5.24 → 4.85 ms.
• fp32: 86.2 → 85.1 ms.

COLMAP matching speed:

• Setup: exhaustive matching without geometric verification on ETH3D DSLR delivery_area + terrains + office (2132 pairs), with the keypoints of the evaluation, on an idle A100.
• ms/pair = total matching time / pairs, including image decoding and the descriptor. Two repeats at 2k, one at 8k.

Matches: identical up to floating-point rounding; 669 instead of 668 of the 2132 pairs have matches.

Tests: feature/ and controllers/ tests pass.