Reuse precomputed distances in TSP evaluation

[er-s-an]

What changed

• build the neighborhood ordering directly from each precomputed distance matrix
• calculate the completed tour cost from the same matrix
• remove the second set of coordinate-distance calculations from evaluation

Why

GetData.generate_instances() already computes a complete distance matrix, but TSPEvaluation recomputed pairwise distances for neighborhood ordering and recomputed every tour edge from coordinates. Reusing the existing matrix removes redundant work while preserving the generated dataset and evaluator callback API.

Fixes #16.

Validation

• python3 -m py_compile llm4ad/task/optimization/tsp_construct/evaluation.py
• compared the new neighborhood order and tour cost with the previous coordinate-based calculations on a fixed instance
• ran a complete small evaluation while making any call to np.linalg.norm fail; evaluation completed with a finite score, confirming that the precomputed matrix is reused throughout