v0.6.0

A breaking rewrite focused on real-time latency, ease of use, maintainability, and strict no_std.

Added

• Infinite and semi-infinite integration limits for the iterative integrators (use f64::INFINITY /
  f64::NEG_INFINITY), via a domain transform. Accurate for convergent, decaying integrands.
• A runnable example for every module under examples/, and BENCHMARKS.md with accuracy and
  latency figures.
• A criterion benchmark suite under benches/, plus deterministic work-count regression tests that
  fail if a latency optimization is reverted.
• CalcError implements core::error::Error and Display.

Changed

• f64-only API. Dropped the num_complex::ComplexFloat generic; transcendental functions now
  come from libm, so the crate is genuinely no_std.
• Errors are a typed CalcError enum instead of &'static str strings.
• Functions and integrands are taken as generic F: Fn(...) (monomorphized) rather than &dyn Fn.
• Removed the redundant runtime "number of integrations" / "derivative order" arguments; these are
  now inferred from the const-generic array lengths.
• Renamed the solver types to FiniteDifferenceSingle/Multi, IterativeSingle/Multi, and
  GaussianSingle/Multi.
• Unified the vector-field calling convention: line and flux integrals now take
  &[&dyn Fn(&[f64; N]) -> f64; N], matching curl and divergence.
• Gaussian quadrature takes the bare integrand; the tabulated weights carry the e^{-x^2} / e^{-x}
  weighting factor.
• Per-solver settings are now a plain-data config field (e.g. solver.config.method = ...) instead
  of getters and setters.
• Renamed the heap feature to alloc.
• Default iterative interval count is now 120 (a multiple of 12) so the Boole and Simpson rules stay
  exact.

Fixed

• Gauss-Hermite and Gauss-Laguerre no longer double-apply the weighting function.
• Gauss-Laguerre multi-fold recursion (it read the Legendre table and failed to decrement the fold).
• The 3D line-integral z-component, and the reversed / equal / NaN limit checks (negative ranges
  such as [-2.0, 1.0] are now accepted).
• The linear and quadratic approximation formulas and their goodness-of-fit metrics.

Removed

• The num-complex dependency, the ComplexFloat generic, and f32 / complex-number support.