NestedLoopJoin::swap_inputs should check number of right partitions before swapping inputs #17358

Describe the bug

NestedLoopJoin left side has to have single partition, there is a check for this:

   fn execute(
        &self,
        partition: usize,
        context: Arc<TaskContext>,
    ) -> Result<SendableRecordBatchStream> {
        if self.left.output_partitioning().partition_count() != 1 {
            return internal_err!(
                "Invalid NestedLoopJoinExec, the output partition count of the left child must be 1,\
                 consider using CoalescePartitionsExec or the EnforceDistribution rule"
            );
        }
...
}

but no check has been performed when join is created with NestedLoopJoinExec::try_new nor this is checked in NestedLoopJoin::swap_inputs.

I'm not sure if it is expected that additional PhysicalOptimizerRule which will insert CoalescePartitionsExec automatically (like EnforceDistribution) or this is a bug in NLJ .

To Reproduce

No response

Expected behavior

I believe this check should be performed when NLJ is created. Delaying check until execute(...) may be a bit too late.

Also, NLJ::swap_inputs should fail to swap left and right input if it will break this rule

Lastly, JoinSelection should not perform join re-order if right input has more than single partition.

Additional context

This issue is experienced in ballista running TPC-DS query 54, apache/datafusion-ballista#1302 (comment)

Ballista performs join re-ordering based on shuffle file information, before stage runs.
At that point all stages have been created so partitioning can't really change. Swapping join inputs and inserting CoalescePartitionsExec wont work for ballista. If this is expected behaviour, ballista should disable this rule

EDIT: I can't find any physical rule which would "fix" left side with CoalescePartitionsExec