Mixing Matters: Evidence Position Bias Across Sequence Mixers in Long-Context Question Answering