Enforce hard evaluation budgets and checkpoint-safe runs

[isty2e]

Summary

This PR makes evaluation budgeting strict by default. Study.run(...) and Study.optimize(...) now charge max_evaluations against reported logical evaluation cost, including local-search inner evaluations, rather than only counting returned records. Callers that want the old outer-record behavior can still pass count_evaluation_cost=False.

It also adds stop_at_checkpoint_boundary=True so CSA runs can return the latest checkpoint-safe state when the budget ends inside an unsafe generation segment. Structured and SciPy local-search kernels now reserve enough budget for later proposals in the same batch before spending extra local-search evaluations.

User impact

This is a behavior change for code that relied on the old default budget accounting. Over-budget reported evaluation cost now raises EvaluationBudgetExhausted instead of silently assimilating an over-budget step. The changelog and local-optimization/checkpointing docs call out the migration path.

Validation

• uv run --python 3.11 --extra test ruff check pyproject.toml src tests
• uv run --python 3.11 --extra test basedpyright src tests
• uv run --python 3.11 --extra test pytest tests -q
• mypy src/variopt still reports the existing scipy/joblib missing-stub errors only; the branch-introduced kernel error is gone.