fix: validate solver inputs and benchmark results

[bardhh]

QP calls with incomplete constraint pairs could silently solve a different problem, and the fast solver could report success for an infinite unconstrained solution. Validate dimensions and constraint pairs across the adapters, including during JIT tracing, and reject non-finite success while preserving existing backend failure codes and error messages.

Benchmark reporting now rejects missing required metrics, unknown comparison/objective metrics, empty experiments, and mismatched batch result counts. Explicitly unavailable measurements remain null in artifacts and N/A in the sweep display instead of becoming zero violations. Sweep objectives and safety decisions require measured, finite values.

Remove obsolete tutorial migration tests and commented solver scaffolding, add active equality/infeasibility coverage, and extend Ruff checks to scripts and optimization/benchmark tests in both CI and pre-commit. Document the benchmark result contract.

Validation:

• Core suite on Python 3.12: 868 passed, 8 skipped; additional status-preservation regressions passed separately.
• MuJoCo suite on Python 3.10: 206 passed, 3 optional oracle tests skipped, 28 high-memory G1 tests excluded.
• Examples and slow tests: 67 passed.
• Targeted solver/benchmark tests passed on Python 3.10 and 3.12.
• Ruff, Black, isort, mypy, pre-commit configuration validation, and diff checks passed.

Compatibility: incomplete benchmark records must now declare required keys. Use None for unavailable measurements; consumers must handle null summaries. Malformed QP inputs now fail explicitly instead of dropping constraints.