Add C++/Python ACO, PSO, GSA, fuzzy, and hybrid optimization

[Rahuldrabit]

Summary

• add a shared C++17 IContinuousOptimizer interface, common result/configuration types, seeded populations, and a heterogeneous solution-transfer pipeline
• add seven PSO families: global-best, ring local-best, constriction, bare-bones, fully informed, quantum-behaved, and binary
• add graph ACO with Ant System, Elitist AS, Rank-Based AS, Ant Colony System, and MAX-MIN AS, plus continuous ACOR
• add continuous Gravitational Search Algorithm (GSA)
• add standalone fuzzy C-means and a reusable zero-order Sugeno controller
• expose the complete suite through both C++ and Python
• add C++/Python examples, benchmarks, documentation, CTest registration, and cross-platform CI

Python API

Python now exposes:

• SearchConfig, ProgressState, ControlSignal, and OptimizationResult
• all PSO and graph ACO variants
• ContinuousAntColonyOptimizer (ACOR) and GravitationalSearchOptimizer
• FuzzyCMeans and fully configurable FuzzyAdaptiveController
• GeneticAlgorithmAdapter and MetaheuristicPipeline
• subclass hooks for custom Python AdaptiveController and ContinuousOptimizer implementations

The bindings release the main GIL during optimization and safely reacquire it for Python fitness/controller callbacks. Smart-holder lifetime management keeps Python-defined controllers and optimizers valid while the C++ pipeline owns them.

User-controlled composition

• the library never selects an algorithm, creates a hybrid, reorders stages, or enables adaptation automatically
• callers choose each optimizer and its exact position by calling MetaheuristicPipeline.add(...)
• all algorithm parameters and budgets are supplied through public configuration structs; defaults are convenience values, not an automatic tuning policy
• adaptive controllers are opt-in, and FuzzyControllerConfig exposes membership thresholds, improvement scale, and every Sugeno consequent
• valid caller-provided control multipliers are forwarded unchanged; invalid/non-finite values are rejected explicitly

Existing GA improvements

• evaluate each mutated offspring exactly once
• avoid constructing bound vectors for every mutation
• support seed solutions for cross-algorithm transfer
• preserve the global best even when elitism is disabled
• report exact evaluation and iteration counts
• validate configurations, seeds, and non-finite fitness values

For the default GA shape (N=50, about 3 elites, crossover rate 0.8), removing pre-mutation crossover evaluations reduces expected offspring objective calls from about 86.4 to 47 per generation (about 45.6%). The exact reduction depends on crossover rate and elite count.

Validation

• Release build and CTest: 10/10 suites passed, including Python bindings
• AddressSanitizer + UndefinedBehaviorSanitizer: 10/10 suites passed
• Python coverage: all 7 PSO variants, all 5 graph ACO variants, ACOR, GSA, FCM, fuzzy control, multithreaded callbacks, hybrid order, and Python extension hooks passed
• standalone Python GA → PSO → ACOR example passed
• wheel build, isolated installation, import, and optimization smoke test passed
• strict -Wall -Wextra -Wpedantic -Wshadow -Wconversion -Wsign-conversion compile for modified headers/examples passed
• Linux, macOS, and Windows Debug/Release CI is configured

Scope note

“Every ACO/PSO type” is not a finite set in the literature. This PR implements the major distinct families listed above and provides extension interfaces for additional C++ or Python optimizers without changing the user-defined pipeline.