Torch-style runtime front-end

This is the public umbrella for the low-level PyTorch-style runtime layer.

The split is intentional:

• Torch.Core defines imperative tensor references, parameters, eager sessions, operation wrappers, compiled scalar/output wrappers, and simple scalar trainers.
• Torch.LinkedSession records the same style of imperative computation into the proved GraphData IR and exposes the theorem connecting compiled runtime backprop to proved graph backprop.
• Torch.Utils contains small demo/training conveniences such as deterministic initializers, small sample builders, and trainer loops.

TorchLean/* builds the higher-level model/program API on top of this layer. So Torch is the low-level session/ref bridge; TorchLean is the nicer user-facing model stack.