Replace hardcoded canonical rule examples with ILPSolver calls

[isPANN]

Summary

• Convert 46 ILP reduction rules from hardcoded SolutionPair to dynamic ILPSolver::new().solve() calls in canonical_rule_example_specs()
• Add z upper bound constraint in MinMaxMulticenter ILP formulation to prevent HiGHS stalling
• Update constraint count assertion in minmaxmulticenter test
• Relax exact-value assertion in IntegralFlowBundles example-db test (solver may return different valid solutions)

Details

ILP<bool> (36 rules): Mechanical replacement — reduce source, solve ILP, extract solution, pass into SolutionPair.

ILP<i32> (10 rules): Same pattern. MinMaxMulticenter additionally needed an explicit z upper bound (max weighted distance) so HiGHS doesn't see [0, 2^31) and hang.

All 3530 tests pass (3383 lib + 66 integration + 81 doc).

Test plan

• cargo build --features example-db,ilp compiles cleanly
• cargo test --features example-db,ilp — all tests pass
• rule_specs_solution_pairs_are_consistent validates all canonical examples

Closes #772

🤖 Generated with Claude Code

[isPANN]

Canonical rule example convention change

After this PR, the standard pattern for canonical_rule_example_specs() in *_ilp.rs files is:

#[cfg(feature = "example-db")]
pub(crate) fn canonical_rule_example_specs() -> Vec<crate::example_db::specs::RuleExampleSpec> {
    vec![crate::example_db::specs::RuleExampleSpec {
        id: "<source>_to_ilp",
        build: || {
            let source = /* construct instance */;
            crate::example_db::specs::rule_example_via_ilp::<_, bool>(source)  // or i32
        },
    }]
}

What changed:

• rule_example_via_ilp<S, V>(source) replaces the old pattern of hardcoding SolutionPair { source_config: vec![...], target_config: vec![...] }
• Internally it does: reduce_to() → ILPSolver::new().solve() → extract_solution() → assemble — in one call, with no double reduction
• The use crate::export::SolutionPair import is no longer needed for ILP rules
• rule_example_with_witness() still exists for non-ILP target rules that need a manually specified solution pair

For new ILP rules: just use rule_example_via_ilp. If the ILP solver hangs, it means the formulation needs tighter variable bounds (see minmaxmulticenter_ilp.rs for an example of adding a z upper bound).