io.py: fix bug in KafkaCD where only the last batch of triggers was cached

This one snuck its way past the unit tests for KafkaCD, and solves the issue:

For KafkaClassifierData, when running the lookback jobs, I've noticed that for the one hour lookback, it'll report that the number of target times is either 0 or 2. Zero if there aren't any target times, but two if there's any. It's strange because when I run the same stream jobs it'll find target times at each 5 second window so this can't be right.

What was going on was that old data being found by the KafkaCD was always being overwritten by the newest item in the queue, and so this issue wasn't obvious in the streaming jobs, but raised some alarms for me when running the longer lookback jobs.