Fix/simulated annealing local search

[athikha]

1. Implemented an optional local-search hook for the Simulated Annealing implementation to allow a small, greedy/hill-climb refinement of candidate solutions before the SA acceptance step.
2. File changed: simulated_annealing.py
3. Added new init args: local_search: Optional[...] and local_search_iters: int.

• Implemented [simple_local_search] — a tiny deterministic hill-climb helper (module-level, with doctest).
• Integrated the local-search call inside [SimulatedAnnealing.optimize] so a candidate can be refined (and its cost returned) before apply the SA acceptance probability.
• Added [_doctest_local_search_benefit] doctest which runs SA with and without a deterministic local search and asserts the local-search-enabled run is at least as good (cost2 <= cost1). The doctest is deterministic (seeded) and passes locally.
• No other files were changed.

Why this change

• Allows combining a fast local search with global SA exploration, which can improve solution quality while keeping the SA control flow intact.
• The change is backwards-compatible: if [local_search] -- is None, behavior is unchanged.