Regressions between addressit parser and pelias parser

Hi,

The pelias parser seems to do an good job, although, I've noticed a few of regressions between the addressit and the pelias parser versions of the API.

Here's a few I've found:

Query: D-Block mini Park, Dhaka District, Dhaka, Bangladesh
addressit version: D-Block mini Park, Dhaka, Bangladesh
pelias parser version: Mini Park, NS, Canada

Query: kim jong yu paradise, Pyongyang City, P´yongyang-si, North Korea
addressit version: kim jong yu paradise, Pyongyang, North Korea
pelias parser version: Yongyang Road, China

Query: Afognak Native Corp, Anchorage, Municipality of Anchorage, Alaska, United States
addressit version: Afognak Native Corp, Anchorage, AK, USA
pelias parser version: municipality of curepipe, Mauritius

Query: City Auto, Dili, East Timor
addressit version: City Auto, Dili, East Timor
pelias parser version: Auto City Speedway, Vienna, MI, USA

When debugging more, I've found that parts of the input string which are not classified by the pelias parser are skipped in the DB queries sent to Elasticsearch, which gives results that do not even belong to the same country.

Start by reproducing the four example queries and comparing addressit with the Pelias parser, then inspect how unclassified input is turned into Elasticsearch queries. Done should mean the reported queries return results in the intended country and that unclassified input does not cause unrelated-country matches.