Add higher_order_derivatives with cached mixed-partial reuse

[5iri]

This PR adds a new higher-order differentiation path that reuses common intermediate derivative terms across orders instead of recomputing them by repeated calls to derivative. It introduces higher_order_derivatives (scalar + array) which computes full order‑k mixed partial tensors and optionally canonicalizes mixed partials (symmetric=true) to maximize reuse.

Performance (opt-in test)

Run: FD_PERF=1 julia --project -e 'using Pkg; Pkg.test()'
Observed speedups on this machine:

• Examples/columnderivative.jl (1 var, order‑2): ~1.0× (expected minimal benefit)
• Examples/hessian.jl: x^2+y^2+z^2 (order‑2): ~3.3×
• Examples/hessian.jl: x*y*z (order‑2): ~3.0×