Memory optimizations

[psathyrella]

• Throw out cached values once you've merged them into irrelevance
  • somewhat nontrivial, since you have to also remove it from all the other processes' caches
• don't give the whole cache to every process
  • probably not worthwhile, it'd introduce all sorts of shenanigans to work out who needs to go where, and then the final process still has to use just as much memory.
• reduce the number of things the master python process keeps track of, or caching more of them on disc (although the latter isn't necessarily a great option, because we already spend a significant fraction of our time reading output files).

[psathyrella]

• rearrange Run() and RunKSet() so you can run all the ksets for each gene at the same time -- then you can only cache one gene's values at a time
• DONE (didn't help) combine trellis, paths, and, uh, that other map that has the same keys ('cause the keys take up a lot of space, and are time-consuming to look through)
• if fragmentation is the reason the memory usage keeps going up, it could be that pre-allocating a map that's larger than you'll need for any query will fix this.