Autograd engine

This is the public umbrella for TorchLean's low-level eager autograd engine.

• Engine.Tape is the tape: a grow-only array of nodes with local VJPs, and the reverse sweep that accumulates their contributions. It is generic in the type of the values it stores.
• Engine.Core is the CPU side: the tape over shape-erased Spec.SomeTensor values and the CPU operations.
• Engine.LibTorch is the CUDA side: the same tape over device buffers, the CUDA operations, FFI kernels, and proof layer native contracts.

Higher-level APIs should usually import NN.Runtime.Autograd.Torch or NN.Runtime.Autograd.Model; this module is for code that works directly at the tape engine boundary.