Garbage collection in ibd_segments, link_ancestors and simplify

[gtsambos]

All three of these methods rely on an object called ancestor_map (C codebase) or A (Python mockups) to store segments that descend from each ancestral node in the tree sequence. In the case of ibd_segments and link_ancestors, these can be a big memory hog (much bigger than the size of the tree sequences themselves). However, the segments corresponding to a given ancestral node are no longer needed once we've processed all of the edges with that node in the 'child' position of the edge table. By pruning away this garbage periodically, we could improve the memory usage of these methods (at the expense of some runtime).

The difficulty is in the freeing of memory -- since the ancestor_map objects have their memory allocated all at once, they also have to be freed all at once. It's difficult to imagine changing this without substantially complicating the existing code.

one alternative is to periodically create a pruned copy of the ancestor_map, and to then replace the original version with the pruned copy. I'm about to push up a PR that demonstrates this.

[jeromekelleher]

One thing you could do is to use the same approach as msprime, and to use something like its object heap which is specifically designed for repeated reuse of the same memory.

It would actually be simpler in many ways to just malloc and free the segments as they are used (that is, free all the segments for a given node once you are done with them). Maybe it's worth mocking this up as a quick test and seeing how it performs (and whether we really do make significant memory gains from it)?