Plan cannot be built when using virtual column and UNION ALL in SQL query

# Affected Version

Version: 0.20.0-iap3 (imply build)

### Description

I ran into this problem when playing with sub-query UNION ALL requests, introduced in druid 0.20.0. It was working fine until I had to add a virtual column (`0 as expense`) to have the same columns in the two SELECT.

The easiest way I found to re-create this problem is the following.

```sql
SELECT COUNT(*)
FROM (
(SELECT 1 as a FROM "test-bug")
UNION ALL
(SELECT 2 as a FROM "test-bug")
)
```

```
Error: Unknown exception
Cannot build plan for query: SELECT count(*) FROM ( (SELECT 1 as a FROM "test-bug") UNION ALL (SELECT 2 as a FROM "test-bug") )
org.apache.druid.java.util.common.ISE
```

Stacktrace:

Note 1: the query works by replacing `SELECT count(*)` by `SELECT a`. And it doesn't work with any expression involving `a` (for eg: `SUM(a)`.
Note 2: The bug is not related to make the union over the same DS. It's just to simply the example. The original bug was found when doing an union between 2 different tables

"test-bug" is a basic data-source created from a kafka-ingestion task.

Note 3: Surprisingly, the same structure works when the union is performed on a system table.

```sql
SELECT COUNT(*)
FROM (
(SELECT 0 as a FROM INFORMATION_SCHEMA.COLUMNS)
UNION ALL
(SELECT -1000 as a FROM INFORMATION_SCHEMA.TABLES)
)
```

Note 4: a real use case that hits this problem is the following:
```sql
SELECT type, SUM(weight) as weight
FROM (
(SELECT type, weight FROM table1)
UNION ALL
(SELECT type, 1 as weight FROM table2) -- weight doesn't exist in table2
)
GROUP BY type
```

Did I touch something not currently supported by druid UNION ALL, or is it a bug in the query transformation?

Thanks

Reproduce the failure using the provided SQL UNION ALL queries with virtual columns and aggregate expressions, comparing them with the working system-table example. Trace the query transformation and planning path associated with these queries; done means the affected COUNT and SUM queries build and execute plans successfully.

Written by the indexing model from the issue text.