Decouple operator statistics propagation from traversal/caching via statistics_from_inputs

[asolimando]

Is your feature request related to a problem or challenge?

The current statistics_with_args / StatisticsArgs design (#21815) embeds cache lookup
and child traversal directly inside each operator's statistics_with_args override via args.compute_child_statistics(...). This means:

• Each operator must be aware of caching mechanics, coupling local propagation logic to the traversal strategy
• Evolving the traversal or cache model requires touching every operator implementation

Describe the solution you'd like

Introduce a stateless statistics_from_inputs method on ExecutionPlan (defaulting to Statistics::new_unknown) that expresses only local propagation logic from pre-computed child statistics:

fn statistics_from_inputs(
    &self,
    input_stats: &[Arc<Statistics>],
    partition: Option<usize>,
) -> Result<Arc<Statistics>> {
    Ok(Arc::new(Statistics::new_unknown(self.schema().as_ref())))
}

The external StatisticsContext owns traversal and cache management, calling statistics_from_inputs after resolving child statistics. statistics_with_args remains the public API and is unchanged.

Benefits

• Non-breaking: statistics_from_inputs has a safe default; statistics_with_args is unchanged
• Operators that override statistics_from_inputs automatically benefit from any future
  improvements to the traversal/caching strategy without code changes
• Operators become easier to test in isolation (no need to construct StatisticsArgs or
  a plan tree)

Describe alternatives you've considered

Keep the current statistics_with_args design as-is. Each operator handles caching via args.compute_child_statistics(...). Works correctly but tightly couples propagation logic to traversal mechanics, making the cache model hard to evolve.

Additional context

Suggested by @2010YOUY01 in this comment as a follow-up for #21815