Last minute fixes

Found a fix for those missing channels in streaming training jobs. I was able to test that this works on setting a lookback, but wasn't able to test on live data since neither detector is locked at the moment. I don't think this will be a problem. The issue was that I was too conservative in repopulating the file cache, which should be repopulated when the cache extent doesn't fully cover the segments in ClassifierData anymore.

I made the assumption that the span of time initially requested by a batch job would fully cover any future segments requested so that this cache would never need to be changed. Worked great for the batch jobs, turns out to be a terrible assumption for streaming jobs. Now I added a few checks to see if/when I need to repopulate the cache.

While doing this, I found 2 things to fix:

• Fixed a typo in --initial-lookback option in idq-streaming_train.
• Remove the use of glue.LIGOTimeGPS in utils.py. This was here because at the time, lal's version of LIGOTimeGPS wasn't pickleable but that had been fixed a while ago, so I'm switching back to lal's version.