Run SortAggregateExec as a native hash aggregate without its input sort #7146

What is the problem the feature request solves?

Spark plans SortAggregateExec when an aggregate buffer is not UnsafeRow-mutable (for example string min/max), and EnsureRequirements puts a SortExec on the grouping keys under it. Comet keeps that sort, so DataFusion runs its ordered aggregate streams on sorted input. The sort exists only because of Spark's hash map limitation, and in Comet it is the expensive part:

• ExternalSorter reserves about 2x each buffered batch, and its spill path gives memory back and has to win it again (External sort can lose reserved spill workspace when memory availability decreases datafusion#24739, fixed on DataFusion main for 56).
• On input sorted on every grouping key, the partial aggregate cannot spill and fails when a reservation is refused. The hash partial emits early instead.

On a job grouping by 87 columns with string min/max, 39 of 40 map tasks failed with Additional allocation failed for OrderedPartialAggregateStream[0], and retries failed in ExternalSorterMerge[0].

Describe the potential solution

When a SortAggregateExec sits on the SortExec for its grouping keys, run the native aggregate over the sort's child:

1. Partial aggregate under a shuffle exchange: drop the sort and report no output ordering.
2. Final aggregate: drop the sort, and keep ordered_by_grouping_keys so the native planner sorts the aggregated output, unless the parent is an exchange.
3. Leave first, last and collect_list on the current path, and gate the change behind a config.

Additional context

Apache Auron does this in AuronConverters.convertSortAggregateExec, unless a parent needs the order. #6936 (hash-mode string min/max under-reports memory) should land first or together. Related: #1994, #4565, #6880.