Drop cuOptCreateRangedProblem workaround for multi-GPU PDLP

[0x17]

• Bug fix follow-up: switch gmscuopt.c back from cuOptCreateRangedProblem to cuOptCreateProblem (row types + RHS)
  • the ranged form was only needed because multi-GPU PDLP without presolve saw 0 constraints ([BUG] Multi-GPU PDLP sees 0 constraints for problems created with cuOptCreateProblem (row types + RHS) NVIDIA/cuopt#2042)
  • fixed upstream in Fix bug #2042 absence of logs on mPDLP error with C API and not using rhs when present NVIDIA/cuopt#2091, first nightly containing it is 26.10.0a278
  • removes the now unused constraint_lower_bounds/constraint_upper_bounds arrays
  • skipping initial solutions for multi-GPU PDLP stays as is
• Verified locally (cu13, single GPU, method 1, num_gpus -1, presolve 0, which still goes through distributed PDLP):
  • trnsport with a278: 5 constraints, optimal
  • same link against pre-fix a276: 0 constraints, solve error
  • gamslib regression with a278: 73/73 match CPLEX
• Not tested on an actual multi-GPU machine