## What changes were proposed in this pull request?

When execute a Python UDF, we buffer the input row into as queue, then pull them out to join with the result from Python UDF. In the case that Python UDF is slow or the input row is too wide, we could ran out of memory because of the queue. Since we can't flush all the buffers (sockets) between JVM and Python process from JVM side, we can't limit the rows in the queue, otherwise it could deadlock.

This PR will manage the memory used by the queue, spill that into disk when there is no enough memory (also release the memory and disk space as soon as possible).

## How was this patch tested?

Added unit tests. Also manually ran a workload with large input row and slow python UDF (with  large broadcast) like this:

```
b = range(1<<24)
add = udf(lambda x: x + len(b), IntegerType())
df = sqlContext.range(1, 1<<26, 1, 4)
print df.select(df.id, lit("adf"*10000).alias("s"), add(df.id).alias("add")).groupBy(length("s")).sum().collect()
```

It ran out of memory (hang because of full GC) before the patch, ran smoothly after the patch.

Author: Davies Liu <davies@databricks.com>

Closes #15089 from davies/spill_udf.