add support for automatic differentiation using Tapenade

[aerorohit]

Summary

• Adds autodiff_tapenade.py integrating Tapenade as an automatic differentiation source
  transformation tool for Compyle-annotated functions.
• Provides three user-facing classes:
  • ForwardGrad — forward-mode (tangent) differentiation
  • ReverseGrad — reverse-mode (adjoint) differentiation
  • ElementwiseGrad — differentiates elementwise kernels, with optional OpenMP parallelism
• The pipeline: transpiles the annotated Python function to C via the c backend, invokes Tapenade to generate the derivative
  source, then compiles both together with pybind11 bindings via Cmodule.
• Extends cimport.py with compile_tapenade_source and get_tpnd_obj_dir helpers for Tapenade-specific build steps.
• Adds two examples: examples/autodiff/billiards.py (physics simulation) and examples/autodiff/nn_mnist.py (neural network
  on MNIST).

Test plan

• tests/test_autodiff_tapenade.py — covers ForwardGrad, ReverseGrad, scalar/vector functions, and
  logistic regression gradients
• tests/test_parallel.py — extended with ElementwiseGrad tests
• Requires Tapenade installed; tests should be skipped gracefully if not present