feat: Add outward-rounded interval determinant signs

[acgetchell]

Summary

Add a small fixed-dimension, outward-rounded interval matrix surface with a
division-free determinant-sign query that returns positive, negative, zero, or
inconclusive evidence without pretending an inconclusive interval is an exact
result.

This is not a request for a generic alternate scalar family or a broad interval
arithmetic package. It is a proof-bearing f64 boundary for the small,
robustness-sensitive reductions already central to la-stack's computational
geometry use case.

Downstream motivation: delaunay

Delaunay currently owns a private F64Interval plus subset-DP determinant
evaluator for the relative-coordinate lifted in-sphere matrix. The entries are
derived by subtraction and squaring, so Matrix::det_direct_with_errbound()
cannot certify the exact-real expression after those operations have already
been rounded to single f64 entries.

This is reusable bounded floating-point mathematics, not Delaunay topology or
predicate classification. Delaunay should continue to assemble its geometric
expressions and interpret signs as Orientation/InSphere; la-stack should
own the outward rounding and small determinant proof. Certified dot-product and
linear-form bounds are deliberately tracked separately so the two APIs can be
implemented and audited independently.

This proposal does not replace exact fallbacks. It centralizes a sound fast
filter and returns inconclusive at the boundary.

Requested contract

The exact API can vary, but it should provide proof-bearing equivalents of:

• construct a point interval for a finite f64;
• enclose an exact-real subtraction of two finite f64 values;
• outward-rounded addition, multiplication, negation, and square;
• inspect lower and upper bounds without allowing callers to create inverted or
  non-finite intervals;
• compute the sign of a small square interval-matrix determinant using a
  division-free algorithm through at least dimension 7;
• distinguish Positive, Negative, Zero, and Inconclusive rather than
  coercing overlap with zero into equality;
• surface range failure explicitly so callers can proceed to exact arithmetic.

The core types should remain fixed-size and stack-backed. A dedicated
Interval/IntervalMatrix<D> or narrowly named bounded-result types are both
acceptable; making Matrix<D> generic over arbitrary scalars is not required.

Numerical requirements

• Every successful interval must contain the exact-real result of the stated
  operation on its binary64 inputs.
• Underflow, overflow, subnormal values, cancellation, and signed zero need
  explicit documented behavior.
• The determinant algorithm must avoid unsound interval division around a zero-
  containing pivot; a Leibniz/subset-DP expansion or another justified
  division-free method is appropriate for these small dimensions.
• Inconclusive evidence must be cheap to detect and must never be published as
  an exact sign.

Acceptance criteria

• Property tests compare interval containment and sign decisions with an
  independent BigRational oracle through the supported dimensions.
• Adversarial tests cover cancellation, exact zero, subnormal products,
  overflow, wide intervals, row swaps, and singular matrices.
• Benchmarks cover conclusive and inconclusive determinant workloads
  representative of downstream geometric filters.
• Documentation keeps the bounded f64 surface distinct from the exact
  feature and explains the expected filtered-exact composition.

Downstream cleanup enabled

After release, Delaunay can remove its local F64Interval and interval
determinant DP while retaining geometric matrix assembly, exact rational
fallback, and semantic predicate enums.