Log GC never runs if MRS is under constant write pressure

In KUDU-716 I ran a large TPCH write benchmark where the writer wrote constantly for the whole life of the tablet server. Thus, the tablet server always flushed and never decided to log GC. So, restart's taking quite a long time as it churns through lots of irrelevant log segments.

We should consider either changing priorities for log GC, or decouple it from the maintenance manager thread (running GC is pretty cheap, maybe doesnt' need to be carefully scheduled/prioritized).