Add lp_pdlp.gms regression test for PDLP column reduction and warm starts

[0x17]

• Add examples/models/regression_tests/lp_pdlp.gms, split out of Sync cuOpt options with NVIDIA/cuopt main (8eb20bd) #25, on a random 3000×4000 LP
  • o1–o3: PDLP column reduction forced on (pdlp_hyper_reduced_matrix 1), automatic via large_nnz_threshold 0 in concurrent mode, and off; all must reach the same optimum
  • o4–o7: PDLP warm starts with PSLP presolve, Papilo presolve and pdlp_hyper_project_initial_primal 0/1; the warm solve (bratio 0) must need fewer iterations than the cold one (bratio 1)
• Currently fails on GPU (RTX A1000, CUDA 13, cuOpt nightly 26.10.0a306 / 8eb20bd): every warm solve needs 1500 PDLP iterations vs. 1100 cold
  • the link sets both initial solutions, and cuOpt logs "Mapped the initial solution into the presolved space" (PSLP) or "Skipping LP presolve" (Papilo)
  • but at warm iteration 0 only the primal is used: primal objective and residual match the cold optimum, dual objective is 0 and dual residual equals the cold start's
  • the same happens with presolve 0; a duals-only warm start is bit-identical to a cold start, and negated duals give exactly the same run as correct ones
  • so the initial dual passed via cuOptSetInitialDualSolution has no effect; the assertion itself is valid, since the cold optimum already meets PDLP's 1e-4 tolerances
• Draft until cuOpt uses the initial dual; o1–o3 and the optimality checks of o4–o7 pass