ONNX export: fused GELU operator in the matcher - #6

Stacked on #5. This PR exports the GELU activations of the matcher as ONNX Runtime's com.microsoft::Gelu operator (exact, erf-based) instead of Div / Erf / Add / Mul / Mul.

Why

• ONNX Runtime doesn't fuse the exported GELU pattern (Div(√2) → Erf → Add(1) → Mul(0.5) → Mul(x)). Each of the 36 GELUs per pair therefore runs as five element-wise kernels over a 4096×512 tensor.
• The exporter swaps the nn.GELU modules of the exported copy of the matcher for a custom op, which the translation table maps to the fused operator (as for the attention). --no_fused_attention disables both, for standard ONNX operators.
• Only the matcher changes; the descriptor model is unchanged.

Results

GPU time per pair: ONNX Runtime 1.23, A100, N = 2048, CUPTI kernel profile, device-resident I/O.

Disabling ONNX Runtime's MatMulAddFusion would save another ~1.4 ms on this setup. That was not adopted, because it depends on cuBLAS's kernel choice; see colmap/colmap#4914.

Parity: python onnx/check.py passes (N = 512 / 2048 / 4096, min_covisibility 0 / 0.8):

• fp16: at most 2 differing matches.
• fp32: at most 1 differing match on CUDA, previously exact. ONNX Runtime's Gelu kernel rounds slightly differently than F.gelu.
• The covisibility score is within 1e-3 of PyTorch.