Use struct instead of named_struct when there are no aliases

[alamb]

Which issue does this PR close?

Closes #9839

Rationale for this change

The names of the columns created by callling struct changed in #9743

Before #9743

select struct(a, b) from t;
+-----------------+
| struct(t.a,t.b) |
+-----------------+
| {c0: 1, c1: 2}  |
| {c0: 3, c1: 4}  |
+-----------------+

After #9743

(note how the column name is different)

select struct(a as field_a, b) from t;
+--------------------------------------------------+
| named_struct(Utf8("field_a"),t.a,Utf8("c1"),t.b) |
+--------------------------------------------------+
| {field_a: 1, c1: 2}                              |
| {field_a: 3, c1: 4}                              |
+--------------------------------------------------+

This change caused #9891 as well as removed test coverage for struct

What changes are included in this PR?

1. Use struct when there are no named fields
2. reverts the test change in [CI] Use alias for table.struct #9894

Are these changes tested?

Yes, by existing tests

Are there any user-facing changes?

The column names are now the same as they were prior to #9894

[alamb]

This reverts the change from #9894

[yyy1000]

I think the values in create_named_struct will be SQLExpr::Named in this PR, so maybe we can also make a change here?
https://github.com/apache/arrow-datafusion/blob/d8d521ac8b90002fa0ba1f91456051a9775ae193/datafusion/sql/src/expr/mod.rs#L610-L631

[yyy1000]

Oh, ignore this. One case is only some values are Named so no change is needed.

[comphead]

thanks @alamb 1 improvement I can see here is we over iterating through values.
I mean 1 iteration is do identify named/not_named and when creating struct/named_struct its another one. prob it can be optimized

[alamb]

thanks @alamb 1 improvement I can see here is we over iterating through values. I mean 1 iteration is do identify named/not_named and when creating struct/named_struct its another one. prob it can be optimized

I agree it does two iterations over the value. I think the current approach basically requires two iterations because if we are doing named_struct we need to add names for each column that don't have a name.

So in other words, struct(1, 2, 3 as other_col) needs to become struct(1 as c1, 2 as c2, 3 as other_col) but we don't know until we see 3 as other_col that we need to add in c1 and c2) 🤔

I would like to spend my optimization budget on bigger things initially if that is ok (specifically #9637)

[alamb]

Thanks again for the reviews 🙏