* Do the string processing to extract the current token once
when autocomplete is started in `source` rather than repeating
it once per potential completion in `match`.
* Store a property for whether we're completing emoji or mention
instead of performing different tests on each step.
This reduces the lag on emoji autocomplete.
(imported from commit 7aa3f762eaf4286608809a09a3cddbbe3dfcabf9)