Investigate det_direct regressions for D=2..4

[acgetchell]

Summary

det_direct() appears to have regressed for D = 2..=4, and det() may also have regressed for D = 4. Investigate the regression, identify the first bad change, and fix the determinant hot paths before v0.4.4.

Current State

• det_direct() is the closed-form determinant path for small fixed dimensions.
• det() should remain competitive while preserving the public error behavior introduced for determinant APIs.
• Recent benchmark results suggest:
  • det_direct() regressed for D = 2, D = 3, and D = 4
  • det() may also have regressed for D = 4
• Existing determinant API refactoring work for generic_const_exprs is tracked separately in Refactor determinant APIs around generic_const_exprs #123 and should not be pulled into this v0.4.4 fix unless it is necessary.

Proposed Changes

• Reproduce the regression against the relevant baseline and current main.
• Determine whether the slowdown comes from:
  • extra fallibility/error-path checks
  • missed inlining or const-folding
  • changed formula structure in the closed-form paths
  • fallback behavior in det() for D = 4
  • benchmark methodology or noise
• Add or update focused benchmarks so the regression remains visible.
• Restore the expected performance without weakening determinant correctness or public error behavior.

Benefits

• Protects the small-dimension determinant hot paths that are central to the crate's scope.
• Keeps v0.4.4 focused on release-train regressions.
• Separates short-term regression repair from longer-term determinant API refactoring.

Implementation Notes

• Compare D = 2, D = 3, and D = 4 independently; the root cause may differ by dimension.
• Check both Matrix::det_direct() and Matrix::det().
• Useful validation commands:
  • just bench-vs-linalg-quick det
  • cargo bench --bench vs_linalg -- det
  • cargo test
  • just check