deprecating `return_type` in favor of `return_fields_from_args`

Having `return_type` does not really reflect the UDF actual nullability.

the following problems exists when UDF wrongly report their nullability:

## Disallow optimization based on the expressions nullability.

For example lets assume I have 2 columns:
1. `col1` which is i64 without nulls
2. `col2` which is i64 with nulls

and the expression `ceil` does not does not implement `return_fields_from_args`, so by default the `nullability` is `true`

can't optimize the following query and

this means that I can can't optimize this query
```
select coalesce(ceil('col1'), 'col2') from tbl;
```

to be this:
```
select ceil('col1') from tbl;
```

because `ceil` says that the nullability is true even though the input expression is not nullable

the same can be for not able to remove the expression `array_remove_all(array, null)` if the array expression report that it is nullable while it is not

## Wrong types

creating the wrong type:

```
> select arrow_typeof(make_array(1));
+------------------------------------------------------------------------------------------------------------------+
| arrow_typeof(make_array(Int64(1))) |
+------------------------------------------------------------------------------------------------------------------+
| List(Field { name: "item", data_type: Int64, nullable: true, dict_id: 0, dict_is_ordered: false, metadata: {} }) |
+------------------------------------------------------------------------------------------------------------------+
1 row(s) fetched.
Elapsed 0.002 seconds.
```

should be nullable `false` but because `make_array` implement `return_type` it does not get that information

Start by locating the UDF implementations that use `return_type`, especially `ceil` and `make_array`, and compare them with `return_fields_from_args`. Trace the nullability and type inference tests, then confirm that affected UDFs derive those fields from their arguments and that the existing SQL examples produce the expected results.

Written by the indexing model from the issue text.