CUDA eager-engine backend

This umbrella collects the CUDA side of TorchLean's eager autograd engine.

The modules here are split by trust boundary by trust boundary:

• Trusted and Buffer expose the opaque FFI buffer type and allocation/copy primitives.
• Kernels, ConvPool, and DGemm declare native CUDA/CPU-stub kernel entrypoints.
• Tape and Ops build the CUDA reverse-mode tape over those buffers.
• Float32Contract and KernelSpec state the proof layer reference contracts for native bits.
• NativeSources documents which C/CUDA files implement the external symbols.

The compiled CUDA binary is a native trust boundary. Lean proves the pure specs and graph-level connections around this boundary, while runtime tests validate the native implementation path.