Onnx-extended

* 189: use onnxruntime==1.19.2 as default, pybind11 2.13.5, MatX 0.8.0
* 187: Fix compilation with GCC>=13 187
* 185: adds custom operator MulMulSigmoid on CUDA
* 184: use onnxruntime==1.18.0 as default
* 181: adds MaskedScatterNDOfShape custom operator
* 175: adds custom operator MulSub and SubMul on CUDA
* 173: adds custom operator AddSharedInput, MulSharedInput on CUDA
* 170: adds custom operator TriMatrix on CUDA
* 169: adds custom operator ReplaceZero on CUDA
* 168: adds custom operator MulSigmoid on CUDA
* 167: adds custom operator Rotary on CUDA
* 166, 178: adds custom operators AddMul, MulAdd on CUDA
* 165: adds custom operators AddAddAdd, MulMulMul on CUDA
* 163: use onnxruntime==1.17.3 as default
* 162: add ScatterNDOfShape implementation on CUDA without atomics
* 159: add AddAdd custom operator on CUDA
* 158: add MulMul custom operator on CUDA
* 157: add ScatterNDOfShape custom operator
* 155: add a function to draw a timeline from a profile
* 154: improves ploting legend for profiling
* 151: refactoring of TreeEnsemble code to make them faster
* 129, 132: support sparse features for TreeEnsemble

* 108: improves command lines documentation, fix an issue in command line stats
* 103: add methods to compute statistics on TreeEnsemble and initializers

* 99: use onnxruntime==1.16.1 as default
* 96: implements a fonction to convert a ModelProto into string (not bytes),
add a function to multiply the number of trees in a TreeEnsemble
* 75: add an implementation of murmurhash3 to validate some options
* 93: validates the wheels in CI
* 89: add a function to merge models and update them if both have different opsets

* 87: update the quantization tools to use a simplified dynamic linear quantization into float 8
* 85: add load_model, save_model to help saving with/without external data
* 82: fixes benchmark on multiple versions of onnxruntime

* 79: update to onnxruntime v1.16.0
* 77: helpers to benchmark a model
* 74: add a function to enumerate all intermediate results with onnxruntime
* 71, 72, 73: add function to analyse a profile produce by onnxruntime
* 68, 69, 70: add CPU implementation for CustomGemmFloat8
* 67: add a function to extract a subgraph of a model
* 59, 60, 61, 62, 63, 65,
66, 68, 69, 70:
add local functions to quantize into float 8, float 16
* 57: add C implementation for DynamicQuantizeLinear (for experimentation)
* 56: add C implementation to cast a float into float 8
* 55, 58: add basic functionality to transform a graph, starts with basic quantization
* 51: fix optmized TreeEnsembleRegressor and adds TreeEnsembleClassifier as custom ops
* 50: add command line store to store intermediate outputs
* 49: add option to save intermediate results in CReferenceEvaluator
* 45: add option cuda-link to setup.py to specify how to link with CUDA library
* 41: implements a custom kernel for RandomForestRegressor easier to optimize
* 34: update to onnxruntime v1.15.1
* 31: implement a custom CUDA kernel (gemm)
* 32: update to onnxruntime v1.15.0
* 27: add a custom kernel with parameters to onnxruntime
* 26: add a custom kernel to onnxruntime
* 24: use Eigen to implement Conv operator
* 23: make pip wheel . work
* 22: rename cmake into _cmake to avoid warnings related to cmake package
* 19: minimal settings to use onnxruntime
* 14: minimal setting to use CUDA
* 8: support for C++ unit test