Scaffolding PyTorch Training Loop

Scaffolds a production-grade PyTorch training loop with deterministic seeding, mixed-precision (AMP) with grad-scaler, gradient clipping, learning-rate scheduling, early stopping on a validation metric, checkpoint save/load with full optimizer + scheduler + scaler state, and run telemetry to TensorBoard or Weights & Biases. Use when starting a new PyTorch training script for a vision / NLP / tabular deep-learning model, when an existing training loop lacks resume-from-checkpoint after a preemption / OOM crash, when training cost is high enough that AMP and LR-scheduling matter, or when results vary run-to-run on the same seed. Refuses to scaffold around sklearn / xgboost …