[SPARK-13830] prefer block manager than direct result for large result

## What changes were proposed in this pull request?

The current RPC can't handle large blocks very well, it's very slow to fetch 100M block (about 1 minute). Once switch to block manager to fetch that, it took about 10 seconds (still could be improved).

## How was this patch tested?

existing unit tests.

Author: Davies Liu <davies@databricks.com>

Closes #11659 from davies/direct_result.