perf: improve LDLT factorization kernel

[acgetchell]

Summary

Follow-up to #137: investigate LDLT factorization performance now that Matrix<D> and Vector<D> carry the finite-entry invariant directly and public solve APIs use the clean solve name.

Current State

The #137 finite-proof API cleanup improved la-stack rows versus the v0.4.2 Criterion baseline, including vector kernels and LU paths. The remaining consistent performance gap is LDLT factorization:

• Ldlt::solve after factorization is often competitive, especially for repeated solves.
• Ldlt factorization itself trails nalgebra/faer in the vs_linalg suite, with the gap becoming more visible as D grows.
• LU factorization and LU solve paths look healthier, so this should remain focused on LDLT rather than revisiting the finite-proof API.

Proposed Changes

Investigate the LDLT factorization kernel for avoidable overhead while preserving the current correctness and API invariants:

• Keep Matrix<D> / Vector<D> finite by construction.
• Do not expose public unchecked constructors or public FiniteMatrix / FiniteVector types.
• Do not reintroduce solve_vec; public factorization solves should remain solve.
• Inspect the LDLT update loops, symmetry handling, and finite internal wrapper path for redundant work.
• Prefer improvements that preserve stack allocation and const-generic fixed dimensions.

Benefits

• Improves the main remaining factorization performance gap after Investigate checked vector kernel performance for v0.4.3 #137.
• Keeps the clean parse-don't-validate API model intact.
• Gives future performance work a narrow target instead of blurring API cleanup and algorithm tuning.

Implementation Notes

• Use benches/vs_linalg.rs as the primary performance signal.
• Compare against the current post-Investigate checked vector kernel performance for v0.4.3 #137 baseline and optionally v0.4.2 for historical context.
• Prioritize D=2 through D=16; D=32 and D=64 are informative but outside the crate's core performance target.
• Keep error behavior unchanged: non-finite raw input is rejected at construction, while LDLT reports computed non-finite intermediates and singular/not-positive-semidefinite conditions as typed LaError variants.

Related: #137