Feature/numba integration

[lenarddome]

Summary

Makes model fitting in cpm much faster, with the public API unchanged. numba is an optional extra (pip install "cpm-toolbox[numba]"); without it, cpm runs the same code as plain Python.

• The built-in applications (RLRW, HybridMBMF, PTSM, PTSM1992, PTSM2025) compute all trials of a participant at once. Fitting them is 90-220× faster with numba and 11-23× faster without, with the same results as main.
• New: cpm.generators.SessionWrapper, for your own models that compute all trials at once, and cpm.models.kernels, the formulas of the cpm.models classes as functions numba can compile.
• Faster priors, per-trial Wrapper runs, losses and meta-d', with identical results.
• Fixes EmpiricalBayes and VariationalBayes discarding the updated population priors after the first evaluation. Hierarchical results change. numpy.random.seed now also reproduces their later chains.
• Adds a test workflow: Linux on Python 3.11-3.14, Windows and macOS on 3.14, each with and without numba, plus runs on the lowest allowed versions and on pre-releases.
• Adds a News section to the docs, and how-to guides on numba and on speeding up your own model.

Other behaviour changes are listed in CHANGELOG.md under 0.26.0.dev0.

Testing

• The outputs of the former per-trial applications are stored in the tests, and both backends must reproduce them exactly.
• The full test suite passes with numba, without it, and with NUMBA_DISABLE_JIT=1. The docs build passes with warnings treated as errors.
• This is the first run of the new test workflow. The lowest-versions job is untested; I'll raise any lower bound that fails.

🤖 Generated with Claude Code