api: add finite Matrix and Vector proof types

[acgetchell]

Summary

Add refined finite-entry matrix/vector types so callers can parse raw Matrix<D> / Vector<D> values once, then run numerical algorithms without repeatedly rediscovering that stored entries are finite.

Current State

Matrix<D> and Vector<D> intentionally accept raw f64 storage, including NaN and infinity, so boundary tests and fallible computations can report structured LaError::NonFinite coordinates. That keeps invalid input representable at the boundary, but hot algorithms must repeatedly validate stored entries before computation.

Proposed Changes

• Add proof-carrying finite-entry types, likely FiniteMatrix<D> and FiniteVector<D>, with fallible constructors from raw arrays and raw Matrix<D> / Vector<D> values.
• Keep raw Matrix<D> / Vector<D> useful at input, fixture, and deserialization boundaries.
• Add infallible or narrower-fallible computation paths on finite types where non-finite stored-entry validation is no longer needed.
• Decide which existing APIs should remain raw-boundary APIs and which should delegate to finite parsed forms.
• Update prelude exports, docs, examples, tests, and benchmarks accordingly.

Benefits

• Makes the finite-entry invariant explicit in the type system.
• Reduces repeated validation in hot paths after parsing.
• Gives downstream callers an ergonomic parse-don't-validate path without losing raw boundary diagnostics.

Implementation Notes

This is a public API design change and may be breaking if existing methods are moved or retargeted around refined inputs. It should land after the broader parse-don't-validate audit work and before adding Semgrep guardrails for examples, benches, and doctests so the examples reflect the settled API.