feat: report basis dimension alongside LU fill and density

[rasros]

Summary

• Adds lpLuMaxDim, the basis dimension at which lpLuMaxDensity was last observed, alongside the existing lpLuMaxFill/lpLuMaxDensity LP stats. Density alone hides whether a dense basis was tiny or large.
• Threads through RevisedSimplex.maxLuDim -> FloatLpResult.luMaxDim -> ComponentLpSolver -> LpBounding.observeLuFill -> LpStats/LpStatsSink, and is printed by CliStats (-s) the same way the other two already are.
• Tracked as its own running max (LpStats.luMaxDim), independent of luMaxDensity — a single running max can't also report the dimension at which the density max occurred across a whole search, so size and density are separate signals rather than a pinned pair.

Why

Motivated by #2006: before building a dense-basis LU fallback, we need to know whether the corpus ever produces small-and-still-hard bases (many repeated factorizations, unresolved search) versus large-and-sparse ones. Run over real MIPLIB2017/MiniZinc/XCSP3 samples, this already found the motivating shape (gr4x6: dimension 34, 10,340 solves, unproved) and confirmed large bases stay sparse regardless of difficulty (nag: dimension 5,796, 0.02% density).

Test plan

• ./gradlew check lintDocs
• :klause:updateKotlinAbi run to reflect the new public LpStats.luMaxDim field
• Manually verified via klause-cli -s and klause-bench solve against real MIPLIB2017/mzn-bench/xcsp3-cop samples