feat: change approx percentile/median UDFs to return floats

[theirix]

Which issue does this PR close?

• Closes Refactor away usage of NUMERICS/INTEGERS in datafusion/expr-common/src/type_coercion/aggregates.rs #18092.

Rationale for this change

1. Migrating to the modern TypeSignature API: 264030c/datafusion/expr-common/src/signature.rs

2. Coercing types of approx_percentile_cont, approx_percentile_cont_with_weight, approx_median to floats. It matches PostgreSQL, DuckDB, and ClickHouse behaviour, except for Spark.

What changes are included in this PR?

• Port remaining UDFs (approx_percentile_cont, approx_percentile_cont_with_weight, approx_median, stub functions) to signature APIs
• Deprecate INTEGERS and NUMERICS arrays in favour of using the TypeSignature API
• They are not removed yet, but marked as deprecated to avoid breaking downstream
• Fix up a SLT for approx_percentile_cont, approx_median to make sure it returns a float

Are these changes tested?

• Tests are passing
• Updated tests to expect floats in return types

Are there any user-facing changes?

[Jefffrey]

Thanks for taking this up. Some of my concerns which held me off on this effort was the impact on error messaging (though I recall there was a recent PR to clean this up #20605) and if this deals with null type inputs correctly now (see #19458)

Also it seems some of the return types are changing now, I think we should call this out?

[theirix]

Thanks for taking this up. Some of my concerns which held me off on this effort was the impact on error messaging (though I recall there was a recent PR to clean this up #20605) and if this deals with null type inputs correctly now (see #19458)

Also it seems some of the return types are changing now, I think we should call this out?

Thank you for the review!

Added more tests for null - seems like it's consistent. The return value of these UDFs was float before (it is seen in UDF doc and also in other SQL engines), so it's reasonable to expect floats in tests too. Do you mean this change of types?

[Jefffrey]

Added more tests for null - seems like it's consistent. The return value of these UDFs was float before (it is seen in UDF doc and also in other SQL engines), so it's reasonable to expect floats in tests too. Do you mean this change of types?

Yes, just referring to how it seems we used to return the input type but now we'll always coerce to float. We can see the return type changing in the SLTs, so would be good to highlight this in the title + PR body (since now we're going beyond a simple refactor)

[Jefffrey]

Since we're now coercing to floats, we can remove some of the implementation code that handles integer types

[theirix]

Agree, removed some code. It also affected approx_median, piggybacking on approx_percentile_cont - updated signature and tests as well.

Clarified the scope of this PR in description - it changes more than expected

[Jefffrey]

Makes sense to change these to return floats since they are meant to be continuous instead of discrete 👍

[theirix]

Yes, it makes sense. Thank you for the review!

[Jefffrey]

With the output type changing, I wonder if that is a significant enough change to warrant an entry in the upgrade guide 🤔