Eager Trainer

Record each forward pass on a fresh tape. CUDA updates keep gradients and optimizer moments on the device; only requested losses and state reads cross back to the host.

def Runtime.Autograd.Torch.Internal.eagerScalarTrainer {α δ : Type} [TorchLean.Storage α] [TorchLean.Storage δ] [Context α] [TensorTransfer α] {paramShapes inputShapes dataInputShapes : List Spec.Shape} (options : Config) (parameters : ParamList α paramShapes) (validateDataInputsIO : TorchLean.TensorPack δ dataInputShapes → IO Unit) (loss : ScalarLoss α δ paramShapes inputShapes dataInputShapes) : IO (ScalarTrainer α δ paramShapes inputShapes dataInputShapes)

Construct the eager CPU or CUDA backend for a scalar trainer.