Use logical null count in case_when_with_expr

[pepijnve]

Which issue does this PR close?

• None, followup to Optimize merging of partial case expression results #18152

Rationale for this change

case_when_with_expr has a code path to handle null values early on in the evaluation. It determines the presence of nulls using value_array.null_count() > 0. For nested arrays this may not be correct. logical_null_count() should be used instead.

What changes are included in this PR?

Check logical_null_count instead of null_count.

Are these changes tested?

Additional tests added to ensure the nested array case is tested. The test already passed before this change, but was a bit less efficient since the null values were tested for equality against each possible when value.

Are there any user-facing changes?

No

[comphead]

Thanks @pepijnve, I ran those tests on main now and their passed 🤔
I would expect them fail without this PR change?

[pepijnve]

Thanks @pepijnve, I ran those tests on main now and their passed 🤔 I would expect them fail without this PR change?

No, that's expected. The code was computing the correct result in spite of this.

What was happens on main is that at https://github.com/apache/datafusion/pull/18872/files#diff-ac23ff0fe78acd71875341026dd5907736e3e3f49e2c398a69e6b33cb6394ae8L869 the null count is 0 so the null handling code path is not taken. The result is that the rows which have an effective null value are tested for equality against each when value. That's pointless since it will be falsy each time.

With this change the null handling code path is taken and the null values get filtered out before reaching the when branch handling.

[alamb]

Thanks @pepijnve -- I went over the tests carefully and it looks good to me

[comphead]

Thanks @pepijnve for the explanation, makes a lot of sense