Add a HistFactory-style likelihood benchmark generator

[guitargeek]

benchmark/histfactory.py generates a scaled-down proxy of the code RooFit emits for a real 1470-parameter ATLAS likelihood, compiles it with the clad plugin, and runs it, printing the compilation time in the same table format as the benchmark executable's primal/gradient/hessian timings. The .cpp is a transient artifact in a temp directory (kept with --cpp); the compiler is read from the clad build tree's CMakeCache, so histfactory.py --clad-build <dir> is all that is needed to compare clad revisions.

The generated model reproduces the statement patterns of the full likelihood at 1/16 scale: 91 parameters, 4 channels with straight-line per-sample yields (flexibleInterp over nuisance subsets, product factors, a lumi parameter), bin loops with uniformBinNumber shape lookups feeding binned nll terms, and a constraint block (gaussian/gaussianIntegral and poisson/poissonIntegral ratios into constraintSum). The driver validates the gradient against central differences of the primal and sampled hessian rows against central differences of the gradient. Measured against the full benchmark, the mini preserves the relative cost profile (gradient ~7x primal vs 12x; hessian-vector product ~2.2x gradient vs 2.9x) while compiling in seconds instead of minutes.

benchmark/HistFactoryMathFuncs.h carries the RooFit::Detail::MathFuncs subset the model calls, with TMath/ROOT::Math replaced by std:: equivalents, the Cephes incomplete gamma behind poissonIntegral, and its custom clad derivatives copied verbatim from ROOT's Math/CladDerivator.h (kept verbatim on purpose, including the Cephes do-while loops and the ROOT formatting, which clang-format off/on markers protect so the file stays diffable against the originals).
constraintSum, the one function carrying #pragma clad checkpoint loop, is emitted into the generated main file instead: clad currently mis-attributes a checkpoint pragma living in an included header, because the planner selects pragmas by raw SourceLocation order, which does not match translation-unit order across files.

The benchmark registers as the CTest test clad-HistFactory (labels "benchmark;short", RUN_SERIAL) under CLAD_ENABLE_BENCHMARKS. Because one of its metrics is the compilation time of the generated code, it compiles at test time via the script instead of going through CB_ADD_GBENCHMARK; CMake passes the tested clang explicitly through the new --compiler option, since the CMakeCache fallback would pick the host compiler on builds configured with gcc.

Output in the CI looks as follows:

[7](https://github.com/vgvassilev/clad/actions/runs/32480829289/job/96766612084?pr=2002#step:30:628)
      Start  6: clad-HistFactory
6: Test command: /usr/bin/python3.12 "/home/runner/work/clad/clad/benchmark/histfactory.py" "--clad-build" "/home/runner/work/clad/clad/obj" "--compiler" "/usr/lib/llvm-20/bin/clang++"
6: Working Directory: /home/runner/work/clad/clad/obj/benchmark
6: Test timeout computed to be: 1200
6: clad build    : /home/runner/work/clad/clad/obj
6: compilation   :     22.867 s
6: n params      : 91
6: function value: 4413.41174684329
6: primal        :      0.023 ms/call
6: gradient      :      0.184 ms/call  (8.1x primal)
6: hessian       :     37.619 ms/call  (204.9x gradient)
6: gradient vs numerical: worst rel. deviation 1.15e-06 (at index 10), 0 outside tolerance
6: hessian vs numerical (8 sampled rows): worst rel. deviation 5.62e-10 (at [18,39]), 0 outside tolerance
6: hessian asymmetry: worst rel. deviation 9.37e-16
6: validation passed
 6/11 Test  #6: clad-HistFactory .................   Passed   23.84 sec

[github-actions]

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 10 out of 262. Check the log or trigger a new build to see more.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[github-actions]

clang-tidy made some suggestions

There were too many comments to post at once. Showing the first 10 out of 22. Check the log or trigger a new build to see more.

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"