v0.7.0

A feature release adding automatic differentiation, a linear-algebra core, nonlinear
optimization, and root finding, while staying strict no_std.

Added

• Automatic differentiation. New scalar module with dual, hyper-dual, and Jet
  number types built on a shared numeric-traits abstraction. Autodiff is now the
  default backend for derivatives, gradients, Jacobians, and Hessians, with finite
  differences still available.
• Linear algebra. Fixed-size, const-generic Vector and Matrix types plus
  small-matrix solves: inverse, LU, Cholesky, QR, and SVD.
• Optimization. A nonlinear least-squares optimization module with Gauss-Newton,
  Levenberg-Marquardt, and trust-region solvers.
• Root finding. A root_finding module with bisection, Newton's method, and Newton
  for systems of equations.
• Pairwise summation utility for numerically stable sums, applied across iterative
  integration and approximation.
• An embedded curve-fitting example and an embedded feature marker that keeps the
  firmware example off host builds.
• A README for every source module, and benchmark suites split by area (calculus,
  linear algebra, optimization, root finding).

Changed

• Autodiff replaces finite differences as the default differentiation method.
• Public integration tests moved from src/**/test.rs into top-level tests/, and
  BENCHMARKS.md was split into per-area files under benches/.