This is helpful if such a test depends on input files that are created
by another test.

Add some `constexpr` info in input tensor structure to the generated
code, inside the same namespace as the session struct.

It looks for example like this:
```c++
constexpr std::array<SingleDim, 1> dim_start{SingleDim{1}};
constexpr std::array<SingleDim, 1> dim_limit{SingleDim{1}};
constexpr std::array<SingleDim, 1> dim_delta{SingleDim{1}};

constexpr std::array<TensorDims, 3> inputTensorDims{
   makeDims(dim_start),
   makeDims(dim_limit),
   makeDims(dim_delta)
};
```

This is helpful for validating inputs when using the SOFIE-generated
code without knowing much about about the original ONNX file, and is the
case for the use in RooFit.

Introduce `RooONNXFunction`, a new `RooAbsReal` implementation that enables
native inference of ONNX models within RooFit. The class loads ONNX graphs,
JIT-compiles them via TMVA SOFIE at runtime, and evaluates them efficiently
with support for automatic differentiation through Clad.

Key features:

* Seamless integration of parametric neural networks into RooFit workflows
* Runtime ONNX-to-C++ code generation via SOFIE (no hard dependency at
  link time, the SOFIE usage is an implementation detail)
* Support for analytic gradients (codegen + Clad)
* Support for serialization to RooWorkspace by embedding the ONNX
  payload as a binary blob

A unit test is also implemented.

This development unlocks new use cases such as neural simulation-based
inference (SBI), likelihood surrogate models, and ML-driven parametric models.

This commit addresses an item on the ROOT plan of work 2026.