feat: Add exact determinant sign via `"exact"` feature flag

[acgetchell]

Summary

Add a det_sign_exact() method to Matrix<D> behind a new "exact" Cargo feature flag, using num-bigint/num-rational for exact rational arithmetic. This keeps the default crate zero-dependency.

Motivation: The delaunay crate needs adaptive precision geometric predicates to fix acgetchell/delaunay#228 — 3D bulk Delaunay construction fails at scale because f64 determinants return wrong signs for near-degenerate configurations. The fix requires a fast f64 filter (existing det()) backed by an exact fallback (det_sign_exact()) for the ambiguous cases.

Proposed Changes

1. Add "exact" feature with optional dependencies

Cargo.toml:

[dependencies]
num-bigint = { version = "0.4", optional = true }
num-rational = { version = "0.4", optional = true }

[features]
exact = ["num-bigint", "num-rational"]

2. Implement det_sign_exact() on Matrix<D>

New file src/exact.rs (only compiled with "exact" feature):

• fn f64_to_bigrational(x: f64) -> BigRational — exact conversion (every f64 = m × 2^e, so this is lossless)
• fn det_sign_exact_impl<const D: usize>(m: &Matrix<D>) -> i8 — Bareiss algorithm (fraction-free Gaussian elimination) in exact BigRational arithmetic, O(D³) operations
• Public method: Matrix<D>::det_sign_exact(&self) -> i8 returning -1, 0, or +1

3. Version bump to v0.2.0

New public API behind a feature flag = minor version bump per semver.

Design Notes

• Why BigRational? Every f64 is exactly representable as a rational number (m × 2^e). Determinant computation via Bareiss in BigRational gives the exact sign with no rounding. For the target use case (matrices up to 7×7 for 5D Delaunay), the cost is ~343 BigRational multiplications — trivially fast.
• Why a feature flag? la-stack currently has zero runtime dependencies. Users who only need f64 operations should not pay for num-bigint/num-rational in their dependency tree.
• Why Bareiss over LU? Bareiss avoids division until the final step, keeping intermediate values as integers (when inputs are integer-representable), which is more efficient for exact arithmetic.

Testing

• Proptest: det_sign_exact() agrees with det().signum() for well-conditioned random matrices (D=2–5)
• Known matrices: exact sign correct for known singular matrices (det=0), near-singular matrices, identity matrices, and Hilbert-like ill-conditioned matrices
• Edge cases: 1×1 matrices, zero matrices, matrices with subnormal entries
• Feature gating: verify the crate compiles without the "exact" feature (no regression)

References

• Bareiss, E.H. (1968). "Sylvester's Identity and Multistep Integer-Preserving Gaussian Elimination"
• Shewchuk, J.R. (1997). "Adaptive Precision Floating-Point Arithmetic and Fast Robust Geometric Predicates"

Downstream

This is a dependency for acgetchell/delaunay#228 — the delaunay crate will use la-stack = { version = "0.2", features = ["exact"] } as part of its new AdaptiveKernel.