AWP: Activation-Aware Weight Pruning and Quantization with Projected Gradient Descent