Autograd training helpers (umbrella import)

This module is the one-stop umbrella for training scripts and runtime tests. It gathers the small helpers that sit on top of the dynamic autograd tape and keeps the import surface stable:

• Runtime.Autograd.Train.* - general helpers (error tagging, gradient/value access, SGD)
• Runtime.Autograd.Train.Dataset - dataset wrapper and deterministic dataloader
• Runtime.Autograd.Train.Eval - evaluation helpers over datasets/batches
• Runtime.Autograd.Train IO - CSV/NPY loaders for small self-contained datasets
• Runtime.Autograd.Train.Logger - pluggable logging helpers
• Runtime.Autograd.Train.Optim - optimizer and scheduler integration
• Runtime.Autograd.Train.Trainer - higher-level trainer API with logging and metrics
• Runtime.Autograd.Train.TapeM.* - TapeM helpers for params, constants, and batch losses
• tensor loader helpers in Runtime.Autograd.Train - vectorOfList, matrixOfLists, datasetOfPairs

NN.Runtime.Autograd.Train is the focused training-helper umbrella. This file is a broader autograd utility umbrella because it also re-exports NN.Runtime.Autograd.Torch.Utils.