bench: scalar functions over dictionary-encoded columns

[radmirnovii]

Which issue does this PR close?

Refs #20935 and #19458.

Rationale for this change

Scalar functions over dictionary-encoded columns have no benchmark coverage:
neither the hand-written dictionary arms (e.g. reverse's, from #23930) nor
the cast away from the encoding that every function without preservation pays
(e.g. encode). This lands the baseline first, so follow-up work generalizing
the arms shows its delta on stable benchmark ids.

What changes are included in this PR?

One criterion benchmark:

• reverse over a dictionary of its own per batch (cold) and one shared
  across batches the way a Parquet column chunk delivers them (warm) — today
  the two cost the same, since an arm cannot reuse anything across batches;
• cast_away: the dictionary cast to its value type and the function called
  once per row — what any function without encoding preservation pays;
• flat: the same rows with no encoding anywhere, as a bound.

encode cannot be called over a dictionary today, so it has no
dictionary-typed groups yet.

Are these changes tested?

cargo bench -p datafusion-physical-expr --bench scalar_function_dictionary
runs clean.

Are there any user-facing changes?

No.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NysGeXTG5opiJKBApsAe5H