feat: autograd support for angle gradients of rotated box

[rahul-flex]

Autograd support for derivative w.r.t rotation angle for box geometry in the Transformed class. Key changes are:

Derivative Computation:

• Applies the chain rule in two stages:

 Uses adjoint sensitivities at face‑mesh sample points to obtain ∂J/∂X_rot. Then derivative of sample points coordinates wrt rotation matrix elements.

Testing:
Finite‑difference vs Autograd checks for seven representative angles across all three rotation axes (x, y, z).

[yaugenst-flex]

Closing in favor of #2475