arrow

ARROW-11076: [Rust][DataFusion] Refactor usage of right indices in hash join

This applies some refactoring to `build_batch_from_indices` which is supposed to make further changes easier, e.g. solving https://issues.apache.org/jira/browse/ARROW-11030

* This starts handling right (1) batch and left (many) batches differently as for the right batches we can directly use `take` on it. This should be more efficient anyway, and also allows in the future to build the index array directly instead of doing extra copying.
* Use `indices.len()` for the capacity parameter, rather than the number of rows at the left. This is of impact at larger sizes (e.g. SF 100), see: https://github.com/apache/arrow/pull/9036 Rather than estimating it based on previous batches, this does it based on the (known) number of resulting rows.
* Reuse "computed" right indices across multiple columns.
* The refactoring makes it easier to apply changes needed for https://issues.apache.org/jira/browse/ARROW-11030 where we need to remove the n*n work that is done for the build side
* The changes don't have a big impact locally on performance on TPC-H with small scale factor, but I believe it should have a similar effect as https://github.com/apache/arrow/pull/9036 on SF=100 by using `indices.len()` rather than the number of rows in the build side.

FYI @jorgecarleitao @andygrove

Closes #9048 from Dandandan/join_right_refactor

Authored-by: Heres, Daniel <danielheres@gmail.com>
Signed-off-by: Jorge C. Leitao <jorgecarleitao@gmail.com>