feat(gpu): add transactional 2-opt search

[pedrozaz]

Summary

• Add gpu::two_opt_route and gpu::two_opt for host-orchestrated best-improvement search using the existing CUDA move selector.
• Apply each move only after an independent full-route f64 cost recomputation confirms strict improvement; return a move count and measured distance improvement.
• Keep caller routes/solutions unchanged on validation, CUDA, or inconsistent-selection errors.

Scope and limitations

• Route membership is preserved; solution-wide CVRP feasibility remains the caller's check.
• No kernel/PTX or dependency changes. Each nontrivial iteration still creates a context and transfers data; no speedup is claimed.
• Public rustdoc, README, user guide, architecture, testing notes, and Unreleased changelog are updated. No release plan is included.

Validation

• just ci — passed (49 portable library tests, 9 hardware-only tests ignored in CI, 5 benchmark tests, 3 doctests, rustdoc, package, deny).
• cargo test --workspace --no-default-features --locked --quiet — passed.
• cargo +1.88.0 check -p vrp-gpu --all-features --locked — passed.
• just kernel-ci — passed.
• cargo test -p vrp-gpu --features gpu --locked -- --ignored --test-threads=1 — 9 passed on NVIDIA GeForce RTX 5060 Ti, driver 615.71.09, compute capability 12.0 (sm_120).
• cargo test -p vrp-gpu --features gpu --release --locked -- --include-ignored --test-threads=1 — 58 tests and 3 doctests passed on the same hardware.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.