NeoMME: Multimodal encoders trained from scratch with a single Transformer