Investigate Vector dot and norm2_sq performance

[acgetchell]

Summary

Track follow-up performance work for Vector::dot and Vector::norm2_sq / squared norm in v0.4.4.

Recent vs_linalg benchmark results suggest these vector kernels are significantly slower than comparable nalgebra and faer rows, unlike the stronger LDLT/Cholesky solve signal. This should be investigated separately from Matrix::inf_norm so each kernel can be profiled, changed, and reviewed independently.

Current State

• bench-vs-linalg includes comparable dot and norm2_sq rows across D=2-64.
• The current implementation appears to lag nalgebra/faer in those rows.
• Local saved-baseline workflows are documented in docs/BENCHMARKING.md for tight optimization loops.

Proposed Changes

• Re-run focused vs_linalg measurements for dot and norm2_sq.
• Profile the current implementations and identify whether the gap is loop shape, bounds checks, missed vectorization/FMA, Criterion setup, or another effect.
• Consider allocation-free and const-generic-friendly implementation changes that preserve the public API and numerical behavior.
• Add or adjust focused benchmarks/tests only where they improve confidence.

Benefits

• Improves small fixed-dimension vector-kernel performance.
• Keeps v0.4.4 performance work scoped alongside the inf_norm investigation.
• Makes README and release performance claims easier to reason about.

Implementation Notes

• Related: Investigate Matrix::inf_norm performance #154
• Use just performance-local-vs-linalg <current-tag> <baseline-tag> for release-signal checks.
• Use just bench-save-baseline <name> vs_linalg, just bench-vs-linalg-la-stack, and just bench-compare <name> vs_linalg all-benches for local iteration.