feat: add PyTorch structural surrogate

[savbcar]

• generates a six-dimensional Latin Hypercube FEA dataset
• evaluates static and modal responses using CalculiX
• adds a PyTorch Dataset and DataLoader workflow
• implements a multi-output MLP surrogate
• standardizes inputs and outputs using training-only statistics
• separates train, validation and held-out test datasets
• adds validation-driven early stopping
• saves best-model state_dict checkpoints
• evaluates held-out stress, displacement and modal-frequency predictions
• generates FEA-vs-predicted and training-loss figures
• reports prediction-error distributions and engineering metrics