Simplify redundant tolerance tests and vector finiteness scanning

[acgetchell]

Summary

Follow up on the whole-repo Rust simplification review with two conservative cleanup targets that preserve the parse/proof invariant model.

Current State

• LU and LDLT each carry local invalid_tolerance_rejected tests even though those tests now only exercise Tolerance::new; factorization APIs accept a proven Tolerance, not raw f64.
• Vector::try_new and FiniteVector::new each implement their own entry-by-entry finiteness scan.
• Matrix code already uses a shared helper pattern for locating the first non-finite stored cell.

Proposed Changes

• Remove redundant LU/LDLT tolerance-constructor tests if central Tolerance coverage remains complete.
• Add a small private vector helper such as first_non_finite_entry_in(&[f64; D]) -> Option<usize> and use it from both public raw parsing and crate-internal proof wrapping.
• Keep public APIs and unchecked constructors unchanged.

Benefits

• Reduces duplicated tests that no longer cover factorization-specific behavior.
• Reduces drift risk between public vector construction and internal proof-bearing vector validation.
• Keeps the parse-don't-validate / proven DTO model intact.

Implementation Notes

• Do not collapse set/set_checked, get/get_checked, Tolerance::new/LaError::validate_tolerance, or factor wrapper types; those preserve API ergonomics, typed diagnostics, compatibility, or invariant naming.
• Validate with just check and cargo test --all-features.

Related: #137, #138