Dataframe has poor performance when computing on many columns with codegen

Suppose we have a dataframe with many columns (e.g 100 columns), each column is DoubleType.
And we need to compute avg on each column. We will find using dataframe avg will be much slower than using RDD.aggregate.

I observe this issue from this PR: (One pass imputer)
https://github.com/apache/spark/pull/18902

I also write a minimal testing code to reproduce this issue, I use computing sum to reproduce this issue:
https://github.com/apache/spark/compare/master...WeichenXu123:aggr_test2?expand=1

When we compute `sum` on 100 `DoubleType` columns, dataframe avg will be about 3x slower than `RDD.aggregate`, but if we only compute one column, dataframe avg will be much faster than `RDD.aggregate`.

The reason of this issue, should be the defact in dataframe codegen. Codegen will inline everything and generate large code block. When the column number is large (e.g 100 columns), the codegen size will be too large, which cause jvm failed to JIT and fall back to byte code interpretation.
This PR should address this issue:
https://github.com/apache/spark/pull/19082
But we need more performance test against some code in ML after above PR merged, to check whether this issue is actually fixed.

This JIRA used to track this performance issue.