Add classic non‑metaheuristic optimisation methods

[eugenioLR]

The library currently focuses on population‑based and trajectory‑based
metaheuristics. Many practical optimization tasks are better solved—or at
least should be compared—with classical derivative‑free or gradient‑based
methods. Providing a small set of well‑tested classic methods would make
metaheuristic‑designer a more complete optimization toolkit and give
researchers a baseline to compare their metaheuristics against, without
needing to leave the library’s ecosystem.

These methods would be exposed as search strategies (or special operators)
and would reuse the existing ObjectiveFunc, Initializer, and stopping
condition machinery.

• Implement Pattern Search (coordinate‑based directional search)
• Implement Nelder–Mead simplex method
• Implement Trust‑Region Interpolated optimisation
• Implement numerical gradient approximation (finite differences)
• Add optional automatic differentiation via autograd
• Implement a basic gradient‑descent operator (optionally with
  momentum and learning rate schedules)
• Ensure all classic methods share the same Algorithm loop and
  reporting so they can be compared head‑to‑head with metaheuristics

Note: gradient based methods might become their own separate issue or even repository.