Efficient Decoding Methods for Language Models on Encrypted Data