fix(backport): account for grouped aggregate memory

[discord9]

Which issue does this PR close?

• Backports the grouped aggregate memory accounting fix from Fix memory size accounting for grouped median and avg apache/datafusion#23357 (commit 73d5d780).

Rationale for this change

Grouped AVG and MEDIAN accumulators under-report their owned state on the GreptimeDB DataFusion 53.1 fork. Under constrained query memory, this can delay spill or rejection and allow high-cardinality aggregation to exceed the intended memory budget.

What changes are included in this PR?

• Account for grouped AVG vector headers, native sum storage, count storage, and null-state allocation.
• Account for grouped MEDIAN native value storage, inner vector allocations, and the outer vector header/allocation.
• Add focused regression tests that compare the reported size with all owned accumulator state, including AVG null-state allocation and empty/populated MEDIAN state.

No public API or dependency changes are included.

Are these changes tested?

Using the lane sccache configuration:

• cargo fmt --all -- --check
• git diff --check
• cargo test -p datafusion-functions-aggregate --lib test_groups_accumulator_size_accounts_for_owned_state (2 passed, 0 failed)

The pre-fix Greptime harness independently reproduced the accounting gap in two four-attempt runs: AVG reported 512 bytes versus 1072 bytes of owned state, and MEDIAN reported 96 bytes versus 1096 bytes.

Are there any user-facing changes?

Yes. Query memory accounting for grouped AVG and MEDIAN now includes all owned accumulator state, allowing existing memory limits and spill/rejection mechanisms to react at the correct threshold. There are no API changes.