fix: make sure JOIN ON expression is boolean type

[jonahgao]

Which issue does this PR close?

Closes #11414.

Rationale for this change

The query in the issue generates the following plan after some optimizations.

select * from t1 inner join t2 on null right join t3 on true;

    Projection: t1.v1, t2.v1, t3.v1
      Right Join:
        Inner Join:
          Filter: Boolean(true) AND NULL
            TableScan: t1
          TableScan: t2
        TableScan: t3

When the simplify_expressions rule attempts to simplify Boolean(true) AND NULL, it fails because the operand types on both sides are different (DataType::Boolean vs DataType::NULL).

The fix is to cast the NULL type to Boolean type.

What changes are included in this PR?

Are these changes tested?

Yes

Are there any user-facing changes?

No

[jonahgao]

This query is available in PostgreSQL, but not available in Spark.
If we plan to support it, we can directly cast the ON expression to boolean without checking its current type.

[alamb]

I think the error is clear and is a reasonable behavior. If it is important for someone's usecase, perhaps we can add support for it then

[alamb]

I think the structure you have now implemented (in the analyzer) makes it pretty easy to support if desired

[alamb]

Thank you @jonahgao (and @2010YOUY01 for applying SQL lancer!)

[alamb]

I think the same rule applies to other APIs (not just SQL).

What would you think about moving the check to LogicalPlanBuilder

datafusion/datafusion/expr/src/logical_plan/builder.rs

Lines 671 to 686 in 6e63748

	pub fn join_on(
	self,
	right: LogicalPlan,
	join_type: JoinType,
	on_exprs: impl IntoIterator<Item = Expr>,
	) -> Result<Self> {
	let filter = on_exprs.into_iter().reduce(Expr::and);
	self.join_detailed(
	right,
	join_type,
	(Vec::<Column>::new(), Vec::<Column>::new()),
	filter,
	false,
	)
	}

That way this check would be applied to the DataFrame API as well as any custom plans that were created?

[jonahgao]

Done. I re-implemented it in type_coercion to avoid creating a temporary joined schema. Also added a test for DataFrame.

[alamb]

I think the error is clear and is a reasonable behavior. If it is important for someone's usecase, perhaps we can add support for it then

[jonahgao]

It's redundant because the reduce operation will be performed inside the join_on function.

[alamb]

I think the structure you have now implemented (in the analyzer) makes it pretty easy to support if desired