Add a non-blocking snapshot API for available distributed task metrics

[sesteves]

Problem

The public metrics-rewriting API waits for all worker tasks to report metrics. Complete metrics may be unavailable after cancellation, task failure, timeout, or shutdown, even though some reports have already arrived.

There is currently no public way to consume those available metrics. The internal rewriter also assumes every expected task is present.

Requested API

Expose a non-blocking, point-in-time API for available task metrics.

The API should:

• Include every available report.
• Tolerate missing reports and gaps in task numbers.
• Preserve existing metrics formats.
• Indicate whether the snapshot is complete.
• Leave the existing complete-metrics API unchanged.

No protocol or worker-reporting changes should be required.

Context

Partial metrics are useful for diagnostics, monitoring, accounting, and resource management. Consumers should be able to use available information without waiting indefinitely or depending on internal implementation details.

[sesteves]

Hi @gabotechs, now that #739 is fixed in #752, I would like to align on the snapshot API for #685 before implementing it.

I propose a synchronous snapshot_distributed_plan_with_metrics(plan, format) that returns a rewritten plan and an is_complete flag. It would read the available task reports once without waiting and support the existing aggregated and per task formats. is_complete would mean every task in the prepared plan has a terminal report. It would not imply query success.

Missing tasks would have no displayed metrics. The snapshot could use empty reports internally while rewriting, without adding them to the stored reports. The existing API that waits for complete metrics would remain unchanged.

Would this return type work for you? Should it expose the raw task reports too? I propose returning an error if the plan has not yet been prepared, consistent with the current rewrite API.