Make the ILP solver backend configurable

[aliddell]

Summary

ILPSolver chooses its solver backend with a fixed list. The method _solve() tries Preference.Gurobi, then Preference.Scip. A caller cannot change this order. This pull request adds a solver_preference parameter. The caller can now set the backend order. The default does not change.

Why

Some users have no Gurobi license. Gurobi then fails on a large model. The solver falls back to SCIP. SCIP is too slow for large graphs. ilpy >= 0.6.0 adds Preference.CuOpt. cuOpt has a native MILP solver under the Apache-2.0 license. It needs no license file. A configurable backend order lets a user put cuOpt between Gurobi and SCIP. A large model then completes without a Gurobi license.

What changed

• ILPSolver.__init__ gains a keyword solver_preference. It accepts an ilpy Preference, a backend name as a string, or a sequence of either. A string matches a Preference by name, without case. Valid names are "gurobi", "scip", and "cuopt".
• None keeps the current order: Gurobi, then SCIP. The change is backward compatible.
• _solve() iterates the configured list instead of the fixed list.
• The warning message no longer names SCIP. It names the backend that failed. It says it tries the next backend.
• New tests cover the default order, a single value, a sequence, an unknown name, an empty sequence, an invalid type, the fallback loop across all backends, and a solve that selects SCIP.

Notes

• This pull request does not change the ilpy version pin. It stays >= 0.5.1. The name "cuopt" needs ilpy >= 0.6.0. The code raises a clear error if the installed ilpy has no Preference.CuOpt.
• The tests need no GPU and no Gurobi license. The solve test uses SCIP.

Test

pytest src/tracksdata/solvers passes. 35 tests pass with ilpy 0.6.0.

🤖 Generated with Claude Code

[JoOkuma]

@aliddell , LGTM, I'll merge once the CIs pass.

[aliddell]

@JoOkuma are we happy with CI?

[JoOkuma]

@aliddell , yes, thank you for the ping. I forgot about this.