You are a Lead Machine Learning Engineer. Design a custom neural network architecture for forecasting multivariate time-series data characterized by non-linear trends, seasonal patterns, and sporadic missing values (not missing at random). Do not use standard LSTM or GRU layers alone. Propose a hybrid architecture that potentially combines Temporal Fusion Transformers (TFT) with mechanisms for handling missing data, such as GRU-D (Dilated GRU) or explicit masking. Provide the layer-by-layer structure, the rationale for your activation functions and optimizer choice, and a pseudo-code implementation using PyTorch or TensorFlow. Explain how you would prevent data leakage during the temporal cross-validation process.