perf: Drop RowConverter from GroupOrderingPartial

[ctsk]

Which issue does this PR close?

• Closes Improve GroupOrderingPartial performance #14565.

Rationale for this change

Faster is better?

What changes are included in this PR?

Are these changes tested?

Yes, added a new test.

Are there any user-facing changes?

No.

(The input_schema argument of GroupOrderingPartial::try_new is no longer used and can be removed, this is not included in this PR)

[ctsk]

Micro-benchmark results

• group_ordering_$n benchmarks the partial group ordering where the ordering contains $n columns. Each Columns consists of 8192 Int32s.

group_ordering_partial/order_indices_1
                        time:   [16.814 µs 17.340 µs 17.934 µs]
                        change: [-71.063% -70.318% -69.568%] (p = 0.00 < 0.05)
                        Performance has improved.
group_ordering_partial/order_indices_2
                        time:   [23.585 µs 23.907 µs 24.176 µs]
                        change: [-65.486% -64.711% -63.930%] (p = 0.00 < 0.05)
                        Performance has improved.
group_ordering_partial/order_indices_4
                        time:   [28.452 µs 28.603 µs 28.794 µs]
                        change: [-52.826% -52.404% -51.942%] (p = 0.00 < 0.05)
                        Performance has improved.
group_ordering_partial/order_indices_8
                        time:   [34.737 µs 34.828 µs 34.935 µs]
                        change: [-64.151% -63.876% -63.589%] (p = 0.00 < 0.05)
                        Performance has improved.

[2010YOUY01]

Thank you, this looks good to me. Let's get the CI fixed.

https://github.com/apache/datafusion/blob/main/datafusion/core/tests/fuzz_cases/aggregate_fuzz.rs should have good coverage for edge cases in group by with partial ordering, so I think this change is safe

[ctsk]

Thank you for taking the time to review!

[lewiszlw]

Why not remove this arg?

[2010YOUY01]

This way we can avoid API change, though I don't have a strong preference

[alamb]

maybe we can add a comment or something as a follow on pr

[ctsk]

Additionally, the input_schema argument in order::GroupOrdering::try_new will also be unnecessary

datafusion/datafusion/physical-plan/src/aggregates/order/mod.rs

Lines 43 to 58 in 1626c00

	impl GroupOrdering {
	/// Create a `GroupOrdering` for the specified ordering
	pub fn try_new(
	input_schema: &Schema,
	mode: &InputOrderMode,
	ordering: &LexOrdering,
	) -> Result<Self> {
	match mode {
	InputOrderMode::Linear => Ok(GroupOrdering::None),
	InputOrderMode::PartiallySorted(order_indices) => {
	GroupOrderingPartial::try_new(input_schema, order_indices, ordering)
	.map(GroupOrdering::Partial)
	}
	InputOrderMode::Sorted => Ok(GroupOrdering::Full(GroupOrderingFull::new())),
	}
	}

[ctsk]

I've added a TODO comment to the code