### What changes were proposed in this pull request?
Currently, `join()` uses `withPlan(logicalPlan)` for convenient to call some Dataset functions. But it leads to the `dataset_id` inconsistent between the `logicalPlan` and the original `Dataset`(because `withPlan(logicalPlan)` will create a new Dataset with the new id and reset the `dataset_id` with the new id of the `logicalPlan`). As a result, it breaks the rule `DetectAmbiguousSelfJoin`.
In this PR, we propose to drop the usage of `withPlan` but use the `logicalPlan` directly so its `dataset_id` doesn't change.
Besides, this PR also removes related metadata (`DATASET_ID_KEY`, `COL_POS_KEY`) when an `Alias` tries to construct its own metadata. Because the `Alias` is no longer a reference column after converting to an `Attribute`. To achieve that, we add a new field, `deniedMetadataKeys`, to indicate the metadata that needs to be removed.
### Why are the changes needed?
For the query below, it returns the wrong result while it should throws ambiguous self join exception instead:
```scala
val emp1 = Seq[TestData](
TestData(1, "sales"),
TestData(2, "personnel"),
TestData(3, "develop"),
TestData(4, "IT")).toDS()
val emp2 = Seq[TestData](
TestData(1, "sales"),
TestData(2, "personnel"),
TestData(3, "develop")).toDS()
val emp3 = emp1.join(emp2, emp1("key") === emp2("key")).select(emp1("*"))
emp1.join(emp3, emp1.col("key") === emp3.col("key"), "left_outer")
.select(emp1.col("*"), emp3.col("key").as("e2")).show()
// wrong result
+---+---------+---+
|key| value| e2|
+---+---------+---+
| 1| sales| 1|
| 2|personnel| 2|
| 3| develop| 3|
| 4| IT| 4|
+---+---------+---+
```
This PR fixes the wrong behaviour.
### Does this PR introduce _any_ user-facing change?
Yes, users hit the exception instead of the wrong result after this PR.
### How was this patch tested?
Added a new unit test.
Closes#30488 from Ngone51/fix-self-join.
Authored-by: yi.wu <yi.wu@databricks.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
a082f4600b