Frequency-Aware Masked Autoencoders for Multimodal Pretraining on Biosignals