Memorization Capacity for Additive Fine-Tuning with Small ReLU Networks