### What changes were proposed in this pull request?
The `RowBasedKeyValueBatch` has two different implementations depending on whether the aggregation key and value uses only fixed length data types (`FixedLengthRowBasedKeyValueBatch`) or not (`VariableLengthRowBasedKeyValueBatch`).
Before this PR the decision about the used implementation was based on by accessing the schema fields by their name.
But if two fields has the same name and one with variable length and the other with fixed length type (and all the other fields are with fixed length types) a bad decision could be made.
When `FixedLengthRowBasedKeyValueBatch` is chosen but there is a variable length field then an aggregation function could calculate with invalid values. This case is illustrated by the example used in the unit test:
`with T as (select id as a, -id as x from range(3)),
U as (select id as b, cast(id as string) as x from range(3))
select T.x, U.x, min(a) as ma, min(b) as mb from T join U on a=b group by U.x, T.x`
where the 'x' column in the left side of the join is a Long but on the right side is a String.
### Why are the changes needed?
Fixes the issue where duplicate field name aggregation has null values in the dataframe.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Added UT, tested manually on spark shell.
Closes#30788 from yliou/SPARK-33726.
Authored-by: yliou <yliou@berkeley.edu>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
512cacf7c6