Add log-density construction for Stein kernels

[sylvesterkaczmarek]

PR Type

Feature.

Description

Fixes #624.

Add SteinKernel.from_log_density to derive the score with automatic differentiation while retaining access to the supplied density. Density parameters remain visible to JIT, differentiation and PyTree updates. The existing score-function constructor is unchanged.

How Has This Been Tested?

• 17 focused tests passed, covering analytic scores, kernel derivatives, JIT, parameter updates, density gradients and numeric inputs.
• Pre-commit checks passed, including Pyright and Pylint.
• Full unit run (python -m pytest tests/unit -q): 1,751 passed, 19 skipped.
• Final focused run (python -m pytest tests/unit/test_stein_log_density.py -q): 17 passed.
• GPU, notebook and full integration suites were not run.

Breaking changes

None. API documentation and CHANGELOG updated.

[sylvesterkaczmarek]

Rebased onto current main and resolved the changelog conflict while preserving both upstream and #624 entries. The Stein log-density feature remains scoped to three files. All 17 focused tests pass under the current dependency set, the complete pre-commit suite passes (Ruff, Pyright, Pylint, cspell, Pyroma and formatting), and the full current unit suite passes: 1,850 passed, 19 skipped. The PR is now 0 commits behind and mergeable; fresh workflows are awaiting maintainer approval.

[sylvesterkaczmarek]

@tc85324 This remains 0 behind, mergeable and SSH-verified on the exact validated head. The Stein log-density change is still scoped to three files; all 17 focused tests pass, the full unit suite passes 1,850 tests with 19 skipped, and the complete pre-commit suite passed. The only current GitHub-side blocker is maintainer approval of the fork-triggered Actions, which are all sitting at action_required. Ready for review once those checks are allowed to run.