[SPARK-17296][SQL] Simplify parser join processing.

## What changes were proposed in this pull request?
Join processing in the parser relies on the fact that the grammar produces a right nested trees, for instance the parse tree for `select * from a join b join c` is expected to produce a tree similar to `JOIN(a, JOIN(b, c))`. However there are cases in which this (invariant) is violated, like:
```sql
SELECT COUNT(1)
FROM test T1
     CROSS JOIN test T2
     JOIN test T3
      ON T3.col = T1.col
     JOIN test T4
      ON T4.col = T1.col
```
In this case the parser returns a tree in which Joins are located on both the left and the right sides of the parent join node.

This PR introduces a different grammar rule which does not make this assumption. The new rule takes a relation and searches for zero or more joined relations. As a bonus processing is much easier.

## How was this patch tested?
Existing tests and I have added a regression test to the plan parser suite.

Author: Herman van Hovell <hvanhovell@databricks.com>

Closes #14867 from hvanhovell/SPARK-17296.