druid sql - left join woes (inconsistent behaviors)

Trying to look into [11150](https://github.com/apache/druid/issues/11150), I got different results, but the results were very inconsistent. Here are some examples of wrong results. I'm not sure what the pattern is, so I tried different combinations, a laundry list follows.

This was done on 0.20.0. I used wikipedia, and a lookup with keys 0-5 and values 0,1,null, and a table with 10 rows, dimensions 1-10 and metrics 1-10.

wikipedia has 24,433 rows.

starting with just wikipedia:

`select * from wikipedia t1 left outer join (select * from wikipedia where sum_added = 1) t2 on t1.sum_added = t2.sum_added -- returns 1 row, with the right-hand side all null`
but
`select count(*) from wikipedia t1 left outer join (select * from wikipedia where sum_added = 1) t2 on t1.sum_added = t2.sum_added -- returns 423,225`
selecting a couple of columns, instead of *, works:
`select t1.sum_added, t2.sum_added from wikipedia t1 left outer join (select * from wikipedia where sum_added=1) t2
on t1.sum_added=t2.sum_added - returns 24,433 rows`

trying to join with the lookup, joining on the key works:
`select * from wikipedia left outer join lookup.myLookup on sum_added=k -- 24,433 results`
but joining on the value gives an error:
`select * from wikipedia left outer join lookup.myLookup on sum_added=v -- unknown exception, unknown type[class java.util.ArrayList] org.apache.druid.java.util.common.parsers.parseException`

joining to my small "regular" table behaves the same as the joins to subqueries on wikipedia:

`--select * from wikipedia left outer join oneToTen on sum_added=myDim -- one row, right side all null`
`--select count(*) from wikipedia left outer join oneToTen on sum_added=myDim -- 24,433`
joining on the metric gives an error:
`select * from wikipedia left outer join oneToTen on sum_added=myMetric -- unsupported operation / java.lang.UnsupportedException`
naming all the columns, instead of giving \*, also gives 1 result:
`SELECT wikipedia.__time, channel, cityName, comment, "count", countryIsoCode, countryName, diffUrl, flags, isAnonymous, isMinor, isNew, isRobot, isUnpatrolled, metroCode, namespace, page, regionIsoCode, regionName, sum_added, sum_commentLength, sum_deleted, sum_delta, sum_deltaBucket, "user", oneToTen.__time, myDim, myMetric
FROM wikipedia LEFT OUTER JOIN oneToTen ON sum_added = myDim -- 1 result`
Naming just the wikipedia columns, not selecting from oneToTen, gives 24,433 results:
`SELECT wikipedia.__time, channel, cityName, comment, "count", countryIsoCode, countryName, diffUrl, flags, isAnonymous, isMinor, isNew, isRobot, isUnpatrolled, metroCode, namespace, page, regionIsoCode, regionName, sum_added, sum_commentLength, sum_deleted, sum_delta, sum_deltaBucket, "user"
FROM wikipedia LEFT OUTER JOIN oneToTen ON sum_added = myDim -- 24,433 results`
Selecting wikipedia.* give 2,310 results!
`SELECT wikipedia.* from wikipedia join oneToTen on sum_added = myDim - 2,310 results!`

Not sure what's going on here. I'll attach an ingestion spec for oneToTen. myLookup is just a map lookup, with the following values:
```
{
"1": "0",
"2": "1",
"3": "0",
"4": null,
"5": "0"
}
```

[ingest10data.txt](https://github.com/apache/druid/files/6362990/ingest10data.txt)
[ingest10spec.txt](https://github.com/apache/druid/files/6362991/ingest10spec.txt)

Start by reproducing the reported joins on Druid 0.20.0 with the wikipedia datasource, myLookup, and oneToTen using ingest10data.txt and ingest10spec.txt. Compare the results for SELECT *, COUNT(*), explicit columns, lookup joins, and metric joins against the reported row counts and errors. Done means the inconsistent left-join results and failures are explained and corrected with regression coverage.

Written by the indexing model from the issue text.