Investigate using levenshtein() instead of pg_tgrm for user search
One of the major shortcomings with pg_tgrm is that it falls on its face if you're looking for something with less than 3 characters. This is easily demonstrated by trying to search or autocomplete any account owned by @hj
It seems based on preliminary discussion and the talk at FOSDEM that pg_tgrm may be overkill for what we need here and if our goal is to simply find results with a simple typo we could do better with a Levenshtein distance of two.
Let's see how this change performs in both accuracy and query performance.