SynthFM: Training Modality-agnostic Foundation Models for Medical Image Segmentation without Real Medical Data