Bump version to 2.2.0

[huangziwei]

rfest/version.py is the single source — pyproject.toml reads it through
[tool.setuptools.dynamic] — so this is the whole change. A local uv build
produces rfest-2.2.0-py3-none-any.whl and import rfest reports 2.2.0.

Release notes, ready to paste

Everything below is the span from the v2.1.2 tag to master: one README commit
and #32 through #38.

---

Corrected — this release changes results

Any hyperparameters selected by fit() on an evidence-optimized model differ
from 2.1.2. If you have published numbers from ASD, ALD, ARD, Ridge,
fASD or sARD, they were produced with the objective described below.

• The evidence objective used the posterior covariance where the Gaussian
  marginal needs its inverse (Fix evidence: use the posterior precision in the quadratic term #32, Use the posterior precision in the sARD evidence quadratic term #34). Not a constant offset, so it changed
  which hyperparameters won: on a smooth 1D RF the shipped objective was
  monotone in the smoothness scale and preferred the largest value on offer,
  where the exact evidence peaks near the true RF width. Reported by @memming
  in negative_log_evidence: quadratic term uses the posterior covariance where the evidence needs its inverse #31, who also fixed the first two of the three copies.
• The prior covariance jitter was absolute rather than relative to the prior
  scale, and the returned C was not the inverse of the returned C_inv
  (Scale the prior covariance jitter to the prior, and return C with its own inverse #33). Worth up to ~143 nats on ASD at moderate smoothness. For ALD at
  the scales in the test suite, the locality prior was being replaced by ridge
  outright, because a 1e-7 jitter swamps prior variances of 1e-37.
• ARD raised AttributeError on every call (Replace np.product, removed in numpy 2, so ARD runs again #37) — np.product, removed
  in numpy 2.
• Every GLM fit raised TypeError (Build the mle slice boundaries without a nested list #36) — jnp.hstack no longer accepts
  a list as an element.

Changed

• Requires Python 3.11 or newer. The old floor of 3.6 was never real: on
  3.10 the resolver lands on a jax from 2023.
• Dependencies are now declared. pip install rfest previously installed
  none of jax, numpy, scipy, scikit-learn or matplotlib — setup.cfg had no
  install_requires and requirements.txt is not consulted by an install.
• opencv-python, used for contour areas in rfest.check, is the contours
  extra: pip install "rfest[contours]".
• Packaging moved to PEP 621 metadata in pyproject.toml; setup.cfg and
  requirements.txt are gone.

Added

• CI running the suite on 3.11, 3.12 and 3.13, and a formatting gate (Add a test workflow, and move packaging into pyproject.toml #35, Format the tree with ruff, and gate it in CI #38).
• Tests for ARD, sARD and the prior kernels, none of which had any, and a
  test comparing negative_log_evidence against a closed-form Gaussian
  marginal derived independently of the implementation.

---

Two notes before releasing

On the number. 2.2.0 as asked. Worth knowing that this both raises the
Python floor and changes numerical output, which is the usual case for a major
bump — say the word if you would rather this were 3.0.0.

On the order with #39. Publishing the GitHub release is what triggers the
upload, so:

• If Publish to PyPI with a trusted publisher instead of an API token #39 is not merged, 2.2.0 publishes with the existing API token. Fine.
• If Publish to PyPI with a trusted publisher instead of an API token #39 is merged and the PyPI publisher is not yet configured, the release
  builds and then fails to upload.

So either release 2.2.0 first and switch afterwards, or configure PyPI, merge
#39, then release.

No tag in this PR — creating the GitHub release creates v2.2.0 and fires the
publish workflow.