Text this: Approximation-Aware Training for Efficient Neural Network Inference on MRAM Based CiM Architecture