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, typed executable graphs, and simple scalar trainers.
• Torch.Initialization defines deterministic parameter initializers.
• Torch.ScalarTrainer provides packed loss and gradient operations; optimizers update its parameters.

NN.Runtime.Autograd.Model builds the model runtime on top of this layer. Application entrypoints live under NN.API.