Autograd engine

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

• Engine.Core is the pure CPU tape over shape-erased Runtime.AnyTensor values.
• Engine.TapeM is a StateT convenience wrapper around the pure tape.
• Engine.FastKernels provides opt-in runtime kernels for hot CPU/GPU paths.
• Engine.Cuda collects the CUDA float32 tape, FFI kernels, and proof-facing native contracts.

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