FSTSuggester shouldn't OOM on large inputs

Currently the input to FSTSuggester needs to be re-sorted and this is done in-memory. Kind of defeats the purpose of the component since everything else is super-efficient but we don't even get to that part because of OOMs during construction.

Robert suggested using a spill-to-disk and merge sort on-disk. I suggested creating a lucene index and then enumerating terms for automaton construction or taking the automaton directly from the index structure (if it isn't pruned).