[FEA] Determinism for MILP heuristics

[anandhkb]

Scope for 26.02:

• Improve work limits in Dual Simplex
• First working version of phase2 in dual simplex
• Deterministic Parallel Branch and Bound without heuristics
• Deterministic exchanges/sync between Heuristics + Branch and Bound
• Minimal working version
• Tuned for performance

[anandhkb]

@aliceb-nv
Please call out your acceptance criteria for 25.10. Merci!

@afender for viz

[aliceb-nv]

Acceptance criteria:
In 25.10, consecutive solver runs in MIP heuristics mode over the MIPLIB2017 benchmark set performed with a given work limit MUST see their objective value per instance remain within a certain epsilon of the corresponding objective value of previous runs. This epsilon remains to determine (<0.01% rel err?).

How to test: perform N (to determine? 5?) runs back-to-back over MIPLIB2017 and heuristics-only=t, with work_limit=W and seed=SEED. The Linf distance between the objective relative error vectors of each pair of runs must remain < rel_err.

[anandhkb]

@aliceb-nv Your Determinism 26.02 PRs have been merged. What do you advise?