feat: add selective non-destructive grouped state reads

[rkrishn7]

Which issue does this PR close?

• Related upstream issue: Add selective, non-destructive reads for grouped aggregation state apache/datafusion#24602

Rationale for this change

Long-lived grouped aggregation state currently must be drained and rebuilt to read a subset of groups. Preserving reads avoid that full-state round trip.

What changes are included in this PR?

• Adds GroupSelection for all groups or ordered group indices.
• Adds optional preserving read APIs and capability checks to GroupValues and GroupsAccumulator.
• Implements selected reads for built-in group values and grouped accumulators.

Are these changes tested?

Yes. Added tests cover repeated and reordered reads, duplicates, nulls, variable-width values, and updates after reads.

Validated with:

• cargo fmt --all
• cargo clippy --all-targets --all-features -- -D warnings
• Relevant datafusion-expr-common, datafusion-functions-aggregate-common, datafusion-functions-aggregate, and datafusion-physical-plan unit tests

Are there any user-facing changes?

Yes. This adds optional public trait methods and GroupSelection. Existing implementations remain compatible through default methods.

[xudong963]

Thanks for the thorough implementation and coverage. I found two correctness issues in valid preserving-read paths: an empty-selection panic for struct MIN/MAX, and unchecked offset truncation for selected variable-width group values. Requesting changes for these two cases.

[zhuqi-lucas]

Read through the API and the impls. Mostly looks right - the evaluate/state split on the adapter lines up exactly with the two Accumulator contracts, and the byte-map keys path gets the offsets and the null slot right. Few things inline.