feat: evaluate surrogate cost accuracy tradeoff

[savbcar]

• adds an independent 200-design FEA holdout dataset
• reserves fixed validation and held-out test sets
• evaluates nested 50/100/250/500-sample training subsets
• repeats training across multiple random seeds
• reports held-out engineering-unit and percentage errors
• measures static+modal CalculiX runtime
• measures end-to-end PyTorch inference latency
• calculates FEA-to-surrogate speedup
• estimates computational break-even point
• visualizes accuracy versus training-data cost
• documents surrogate domain and benchmarking limitations