LFM2.5-Encoders for Fast Long-Context Inference on CPU