Fix MMAP reference validation tests

[GiggleLiu]

The Segmentation 13 and 14 tests compared assignments against external solver output, even though those assignments are feasible results rather than certified optima. Independently scoring both assignments shows TensorInference's candidates have higher log objectives by 31.5093806324 and 3.1385028268 respectively.

This restores all three cases and checks that the returned assignment reproduces the reported objective and meets the external feasible lower bound. It also adds an independent exhaustive oracle over small MMAP models. The oracle derives expected output variables directly from the query and evidence, verifies the model exposes that order, and brute-force enumerates assignment scores without using TensorInference contraction results. A better external score alone is not treated as proof of global optimality.

This is the first agent-created maintenance PR for the repository, so it also adds the required maintenance badge.

Validation:

• Julia 1.12.4, test/mmap.jl: 202/202 assertions passed (CUDA excluded)
• Independent BigFloat conditioned-component enumeration: cases 12/13/14 agree with TensorInference candidate scores
• git diff --check

Closes #42.
Closes #79.