Report a retried Databricks task once, from its last attempt

[moomindani]

A Databricks run carries one entry per task attempt. extract_failed_task_errors[_async] walked every entry, so the failure list that ends up in a task's error message (operators/databricks.py:159, and the trigger event on the deferrable path) had two problems:

• a task that Databricks retried was reported once per attempt, and get_run_output was called once per attempt as well
• a task that failed on an earlier attempt but succeeded on the retry was still reported as failed

Both go away by keeping only the attempt with the highest attempt_number per task_key before filtering for failures.

The semantics I picked, in case anyone wants to argue for the other one: report the last attempt only. An earlier attempt's error does not describe the outcome of the task, and if the retry succeeded there is no failure to report at all. Deduplicating by task_key while keeping the first entry would fix the duplication but keep reporting recovered tasks as failures.

Verified against a live workspace, with a job whose single task always fails and carries max_retries: 1:

• the API returns two entries for that task — attempt_number 0 and 1, both FAILED, with different task run ids
• after the change extract_failed_task_errors returns exactly one entry, carrying the last attempt's run id

All three new unit tests fail without the change, and the provider's unit suite passes with it.

related: #72304

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 5)

Generated-by: Claude Code (Opus 5) following the guidelines

[moomindani]

@eladkal could you take a look when you have a moment, or point it at whoever is better placed?

A run carries one entry per task attempt, so a retried task is reported once per attempt, and a task that failed on an earlier attempt but succeeded on the retry is still reported as failed. Green, and verified against a live workspace.

Note that #70385 edits the same two lines, so whichever lands second needs a trivial rebase.

---

Drafted-by: Claude Code (Opus 5); reviewed by @moomindani before posting