SummaryMixing: A Linear-Time Complexity Alternative to Self-Attention, to Streaming Speech Recognition with a Streaming and Non-Streaming Conformer Transducer