Feat: infeasibility detection

[LucasBoTang]

Summary

Adds primal/dual infeasibility detection to the main solve loop. The earlier implementation (removed in #71) was unreliable; this is a new integration:

• On every termination evaluation, compute_infeasibility_information() builds ray estimates from the current fixed-point residual, and check_termination_criteria() returns PRIMAL_INFEASIBLE / DUAL_INFEASIBLE when the certificate ratio is at or below the new eps_infeasible (default 1e-14).
• eps_infeasible is exposed to Python as InfeasibleTol, with validation and a README entry.
• The infeasible/unbounded tests are re-enabled with InfeasibleTol=1e-6.
• Version bumped to 0.3.1.

Verification

• pytest test/: 55 passed.

• MIPLIB instances (feasible): no false infeasibility declarations.

• Netlib infeasible set (29 instances):

  presolve	ε	detected	time limit	false optimal
  off	1e-4	12/29	16	1
  off	1e-6, 1e-8	12/29	17	0
  PSLP	1e-4	22/29	6	1
  PSLP	1e-6, 1e-8	22/29	7	0

  Not every instance is caught, but this is a clear improvement over the previous implementation.