GF(2) presolve reduction for enlight_hard.mps

[aliceb-nv]

MIPLIB2017 problems from the enlight* series are usually hard to solve by the regular B&B and heuristics process, but they possess a specific problem structure which can be reduced to a properly constrained system of equations in GF(2). (machine translation does a decent job at translating this blog post)

In particular, each constraint is of the form

With x_i binary variables, b_i 0 or 1, and y_i an integer 'dummy' variable used to represent equality modulo 2.

This system can then be solved by textbook Gaussian elimination in GF(2).

Problems enlight_hard.mps (in the benchmark set) and enlight8.mps are thus solved to optimality during presolve, and enlight4.mps, enlight11.mps and enlight9.mps are proven infeasible.
This in turns improves our benchmark performance by +1 feasible and -0.5% average gap.

Benchmark results:

Main branch 53d6e74:

Regression Baseline Summary (Time Limit: 600s):
  Feasible instances: 221/240
  Average gap: 0.207478
  Average integral: 0.261499

Excel results

PR 182ac18

cuopt 600s: 
feasible 222/240 
gap 0.202463 
integral 0.267925

Excel results

[rg20]

Great example of how building new presolver methods without adding new infrastructure!

[rgsl888prabhu]

/merge