Updating the package, now AD and CUDA capable, still zero heap allocations

[hzarei4]

Summary

This PR refactors and extends the transformation implementation in DataToFunctions.jl.

Main changes

• tuple-based affine and homogeneous transformation internals
• allocation-free in-place CPU transformation paths
• optimized polynomial transformations using EvalMultiPoly.jl
• custom ChainRulesCore.jl reverse rules for affine and polynomial parameters
• substantially improved Zygote gradient performance
• compatibility with CUDA-backed arrays without requiring CUDA.jl as a package dependency
• backend adaptation of interpolation objects using Adapt.jl
• support for gradient-based optimization, including Optim.jl/LBFGS
• expanded tests for numerical correctness, allocations, AD, and CUDA
• updated Documenter.jl documentation and tutorial

Performance

The in-place affine and polynomial CPU paths can run with zero heap allocations after compilation.

Polynomial reverse-mode differentiation now uses a custom VJP (Vector-Jacobian Product) rather than tracing the full pixel-wise computation.

Compatibility

The existing public transformation interfaces are retained, including the legacy compatibility API.