[FEA] Expose basis / variable status on the LP solution (companion to reduced costs)

Is your feature request related to a problem? Please describe.

cuOpt's LP/QP solution exposes dual values and reduced costs, but not basis / variable status (basic vs. nonbasic-at-lower-bound vs. nonbasic-at-upper-bound). Basis status is the standard companion to reduced costs in an LP sensitivity report: it says which variables are at a bound, where a reduced cost is a meaningful "this would enter / change if its coefficient improved by X" signal, versus basic / interior, where the reduced cost is ~0. Without it, code reading get_reduced_cost() can't cleanly tell a genuine near-miss from an interior variable.

It is also the degeneracy signal: a basic variable at a bound flags a degenerate vertex, so a consumer can mark that dual directional rather than quote a non-unique number as precise. That trust-calibration is what matters most on large, degenerate models, and the prerequisite for any honest sensitivity layer.

This surfaced building solver-exact sensitivity/explainability on top of cuOpt (a downstream multi-objective decision layer): the at-bound-vs-basic distinction is load-bearing for interpreting reduced costs, and we currently reconstruct it heuristically from the primal values plus variable bounds.

Describe the solution you'd like

A public accessor on the LP solution, e.g. get_variable_status() / get_constraint_status() (C++), with C API and Python equivalents, returning a public basis-status enum (BASIC / NONBASIC_LOWER / NONBASIC_UPPER / FIXED / FREE, plus NO_BASIS when no basis exists). The data already exists internally: the dual-simplex solver computes a full basis (variable_status_t / vstatus, cpp/src/dual_simplex/initial_basis.hpp). So the work is mainly (a) a public enum, (b) plumbing vstatus from the dual-simplex result through to the public optimization_problem_solution, (c) the getter plus C API, cython bindings, and tests.

Two design points for maintainers:

1. Public enum shape. The internal variable_status_t is {BASIC, NONBASIC_LOWER, NONBASIC_UPPER, NONBASIC_FREE, NONBASIC_FIXED, SUPERBASIC}. The call is whether to expose all of these or fold some (SUPERBASIC, or the FREE/FIXED split) into a leaner public enum.
2. PDLP semantics. For the PDLP path the accessor should return an explicit NO_BASIS status rather than a heuristic reconstruction. Since PDLP frequently wins the concurrent race, "no basis exists" is the common case, and an honest signal there is most of the feature's value: it tells the consumer these duals are tolerance-accurate approximations, not basis-exact.

Describe alternatives you've considered

• Reconstructing status downstream from the primal solution plus bounds (what we do now): brittle near bounds, and impossible to get right at degenerate points.
• Inferring from reduced_cost == 0: unreliable under degeneracy (basic variables can have zero reduced cost, and vice versa).

Additional context

• Companion to the merged skills-doc note on per-problem-type dual support: NVIDIA/cuopt#1393.
• Basis status is the more fundamental of the two gaps. It pairs with ranging (#1395): a clean basis certifies the dual as a marginal rate valid within the current RHS range, and ranging supplies that range.
• Prior art: Gurobi VBasis/CBasis, HiGHS HighsBasis (col_status/row_status).
• Discovered via cuOpt × Frontier integration testing (solver-exact duals in a multi-objective decision layer).

Related: the diet LP duals example (dual values + reduced costs), NVIDIA/cuopt-examples#154, surfaces exactly the reduced costs this status accessor would annotate (which foods are at a bound = near-misses vs basic/interior).