Fix zero-cost edges in the TSP example

[zayenz]

The br17 example reports 87 instead of the published optimum of 39 because it treats every zero in the distance matrix as a missing edge. br17 contains valid off-diagonal zero-cost edges.

Remove that extra filtering and let circuit enforce the tour. The example then reaches 39; the two smaller instances retain costs 2276 and 22.

Add a focused CTest regression that runs the real br17 example and requires cost 39. Include it in the Ninja CI test selection so zero-cost edges cannot be filtered again unnoticed.

Checks:

• configured a Release Ninja build with examples and testing enabled
• built the tsp target
• ran example-tsp-zero-cost-edges successfully (0.37 seconds)
• restored the old filter temporarily and confirmed the regression fails with best cost 87
• restored the fix, rebuilt, and reran the regression successfully
• git diff --check

Closes #151.