Fix bulk_get not_found response handling when revs=all

[nickva]

Since revs=all when doc is not found we return {ok, []} we didn't count that response as a not_found response and success_possible would return false too early. If a node was in mm mode and that was the last response we could end up emitting that as an internal error instead of the proper "not found" result.

[rnewson]

This is compounded if you fetch multiple documents at once where one exists and another does not and when they are in different shard ranges. if you receive the messages for the non-existent document before the ones for the doc that does exist, this code will return an error tuple (for reason 'maintenance_mode') for all the attempted gets.