Improve hstack lowering

[rjzamora]

Description

• This is a follow-up to Expand CSE placeholders during HStack lowering #21796
• This (hopefully) simplifies some code in Implement streaming window functions in cudf-polars #22191

Problem statement: We currently translate HStack nodes with non-pointwise expressions to the equivalent Select node at lowering time. This is because all our non-pointwise Expr-decomposition logic is specific to Select. Before this PR, this translation was skipped whenever the underlying HStack was completely overwriting it's original columns. The problem with this case is that we loose "anchor" columns that tell the Select how to broadcast scalar-aggregation results.

Proposed solution: We add a temporary "anchor" column to the translated HStack so that broadcasting works correctly in the Select node.

Motivation:

• We can handle all over() expression decomposition within Select if we know all non-pointwise HStack operations are lowered to Select anyway.
• We don't "fall back" for other non-over HStack corner cases either.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[Matt711]

Looks good one nitpick. And I think we can write a test here.

[rjzamora]

I think we can write a test here.

So, I'm not entirely sure if a new test makes sense. test_with_columns_scalar_upstream_20981 would fail if this code path was implemented incorrectly. The only reason we didn't get an error before is that we didn't define a msg argument to the _lower_ir_fallback call (so we didn't see a UserWarning failure here). Now that there is no silent fallback path, it needs to "work" for the test to pass.