Investigate Matrix::inf_norm performance

[acgetchell]

Summary

Investigate whether Matrix::inf_norm can be made faster without weakening its finite-value and overflow error behavior.

Current State

The v0.4.3 release-signal comparison against v0.4.2 is broadly positive, so this should not block the v0.4.3 release. Larger-dimension local vs_linalg measurements suggest Matrix::inf_norm is the most interesting remaining leaf-kernel target:

• D=16: about 66.7 ns
• D=32: about 294 ns
• D=64: about 2.15 us

Dot and squared-norm rows also deserve a glance, but inf_norm looks more likely to matter for determinant/error-bound and validation-adjacent paths.

Proposed Changes

• Inspect the current Matrix::inf_norm implementation and call sites for avoidable fallibility, iterator, or row-sum overhead.
• Compare clear indexed loops, helper extraction, and any FMA/accumulation variants that preserve the documented semantics.
• Benchmark focused la_stack_inf_norm rows across D=2, 3, 4, 5, 8, 16, 32, and 64 before and after any change.
• Keep any optimization only if it is consistent and does not obscure the mathematical invariant or typed error behavior.

Benefits

• Improves a hot matrix leaf kernel without expanding crate scope.
• Keeps post-v0.4.3 performance work separate from release preparation.
• Gives future release notes a clean, auditable benchmark story.

Implementation Notes

• Preserve LaError::NonFinite { row, col } source-location behavior for non-finite entries or overflowed row sums.
• Do not trade correctness or diagnostics for speed.
• If the investigation finds no robust improvement, close with benchmark evidence rather than forcing a code change.