Speed up MOI backend extraction

[bernalde]

Summary

• Build sparse QUBOTools forms directly from MOI objective terms instead of first materializing large linear/quadratic dictionaries.
• Add a dense TSP-style constructor benchmark fixture, with a PR-safe default and an opt-in 10,000-variable scale case via QUBOTOOLS_SCALE_BENCHMARKS=true.
• Document QUBOTools.Model(moi_model) as the supported public MOI/ToQUBO materialization path.

Tests

• julia --project=benchmark benchmark/test/runtests.jl - passed.
• julia +1.12 --project=. -e 'import Pkg; Pkg.test()' - passed.
• julia --project=benchmark -e '... benchmark_dense_tsp_constructor_fixture("tsp/cities=100", 100) ...' - passed; 10,000 variables, 1,980,000 quadratic terms, 0.887258424 seconds, 209,003,768 bytes allocated.

Branch Hygiene

• Base branch: main
• Source branch point: origin/main at 089966e
• Stacked status: not stacked
• Prerequisite PRs: none

Closes #101

[github-actions]

Performance Report - main vs. dev

Timings are reported for the batched benchmark operations as minimum (median) ± std, and comparison uses a 15.00% tolerance to reduce false positives on shared runners.

case	main	dev	diff
/constructors/n=128/Model/MOI/bool	686.167 μs (743.771 μs) ± 583.827 μs	273.385 μs (293.935 μs) ± 405.620 μs	improvement (-60.16%)
/constructors/n=128/Model	703.302 μs (748.313 μs) ± 974.580 μs	699.706 μs (736.071 μs) ± 971.912 μs	invariant (-0.51%)
/constructors/n=2048/Model/MOI/bool	29.523 ms (31.194 ms) ± 38.367 ms	9.565 ms (10.429 ms) ± 21.548 ms	improvement (-67.60%)
/constructors/n=2048/Model	29.429 ms (33.687 ms) ± 26.634 ms	29.242 ms (32.882 ms) ± 24.606 ms	invariant (-0.64%)
/constructors/n=384/Model/MOI/bool	2.102 ms (2.462 ms) ± 765.996 μs	822.538 μs (893.394 μs) ± 469.064 μs	improvement (-60.88%)
/constructors/n=384/Model	2.531 ms (2.839 ms) ± 1.624 ms	2.519 ms (2.736 ms) ± 1.771 ms	invariant (-0.51%)
/constructors/tsp/cities=36/Model/MOI/bool	110.200 ms (117.047 ms) ± 99.555 ms	37.175 ms (39.814 ms) ± 51.694 ms	improvement (-66.27%)
/constructors/tsp/cities=36/Model	66.249 ms (71.440 ms) ± 59.286 ms	66.469 ms (71.532 ms) ± 57.396 ms	invariant (0.33%)
/conversions/n=128/form/Dict	189.261 μs (229.440 μs) ± 3.009 ms	181.259 μs (219.951 μs) ± 2.949 ms	invariant (-4.23%)
/conversions/n=128/form/Matrix/Float32	39.578 μs (46.258 μs) ± 169.271 μs	39.378 μs (47.360 μs) ± 210.264 μs	invariant (-0.51%)
/conversions/n=128/form/Matrix	51.536 μs (61.271 μs) ± 2.991 ms	50.535 μs (60.509 μs) ± 2.899 ms	invariant (-1.94%)
/conversions/n=128/form/SparseMatrixCSC	158.736 μs (179.115 μs) ± 215.151 μs	178.605 μs (193.066 μs) ± 251.879 μs	invariant (12.52%)
/conversions/n=128/ising/Matrix	51.476 μs (62.082 μs) ± 2.958 ms	52.428 μs (60.710 μs) ± 2.948 ms	invariant (1.85%)
/conversions/n=128/qubo/Matrix/min	643.263 μs (685.781 μs) ± 7.437 ms	632.307 μs (687.584 μs) ± 7.158 ms	invariant (-1.70%)
/conversions/n=128/qubo/Matrix	335.307 μs (373.253 μs) ± 5.337 ms	339.333 μs (375.306 μs) ± 4.398 ms	invariant (1.20%)
/conversions/n=384/form/Dict	371.901 μs (638.887 μs) ± 3.515 ms	382.217 μs (649.772 μs) ± 3.613 ms	invariant (2.77%)
/conversions/n=384/form/Matrix/Float32	170.312 μs (226.696 μs) ± 11.617 ms	172.325 μs (234.318 μs) ± 10.987 ms	invariant (1.18%)
/conversions/n=384/form/Matrix	273.786 μs (371.220 μs) ± 19.145 ms	279.895 μs (398.375 μs) ± 17.659 ms	invariant (2.23%)
/conversions/n=384/form/SparseMatrixCSC	1.139 ms (1.430 ms) ± 8.349 ms	1.407 ms (1.533 ms) ± 6.045 ms	regression (23.51%)
/conversions/n=384/ising/Matrix	278.974 μs (381.385 μs) ± 19.177 ms	318.021 μs (404.709 μs) ± 18.504 ms	invariant (14.00%)
/conversions/n=384/qubo/Matrix/min	6.231 ms (6.405 ms) ± 39.587 ms	6.255 ms (6.417 ms) ± 34.210 ms	invariant (0.37%)
/conversions/n=384/qubo/Matrix	3.155 ms (3.314 ms) ± 30.317 ms	3.193 ms (3.354 ms) ± 28.902 ms	invariant (1.18%)
/evaluation/n=128/value/dense-form	363.548 μs (396.268 μs) ± 686.730 μs	363.910 μs (394.024 μs) ± 557.745 μs	invariant (0.10%)
/evaluation/n=128/value/dict-form	267.596 μs (268.828 μs) ± 5.611 μs	267.566 μs (268.598 μs) ± 6.946 μs	invariant (-0.01%)
/evaluation/n=128/value/model	52.057 μs (82.492 μs) ± 680.651 μs	51.256 μs (53.769 μs) ± 368.677 μs	invariant (-1.54%)
/evaluation/n=128/value/sparse-form	51.176 μs (82.182 μs) ± 683.550 μs	51.256 μs (62.838 μs) ± 496.000 μs	invariant (0.16%)
/evaluation/n=384/value/dense-form	4.221 ms (4.248 ms) ± 198.270 μs	4.219 ms (4.287 ms) ± 203.352 μs	invariant (-0.06%)
/evaluation/n=384/value/dict-form	501.173 μs (509.396 μs) ± 9.242 μs	501.864 μs (509.866 μs) ± 9.508 μs	invariant (0.14%)
/evaluation/n=384/value/model	172.486 μs (181.109 μs) ± 218.221 μs	175.880 μs (185.105 μs) ± 181.197 μs	invariant (1.97%)
/evaluation/n=384/value/sparse-form	171.975 μs (178.825 μs) ± 230.640 μs	175.160 μs (185.214 μs) ± 205.549 μs	invariant (1.85%)

[bernalde]

Blocking issues: none.

Nonblocking issues: none.

Questions: none.

Tests run and outcomes:

• Verified local merge-base diff against GitHub-reported totals: 6 files, +411, -64.
• julia --project=benchmark benchmark/test/runtests.jl passed: Benchmark Fixtures 17/17, Benchmark Suites 8/8.
• julia +1.12 --project=. -e 'import Pkg; Pkg.test()' passed: 1111/1111.
• 10,000-variable dense TSP extraction check passed: 1,980,000 quadratic terms, 1.152567399 seconds, 209,034,792 bytes allocated.
• gh pr checks 103 reports all checks passing.

Merge-readiness: code is merge-ready from this review. The PR is authored by my GitHub account, so this is submitted as COMMENT; a formal approval must come from another maintainer.