[autorevert] fix query sorting (#7043)

Current sorting uses the workflow dispatch time, what does not match the
order for commit sequence. The correct approach is to sort by merge
timestamp for all workflows.

This was causing errors in the detection logic, as it was mixing the
order of jobs for commit evaluation, detecting rules where it should
not.

```
==================================================
SUMMARY STATISTICS
==================================================
Workflow(s): Lint, trunk, pull, inductor, linux-binary-manywheel
Timeframe: 4380 hours
Commits checked: 33873
Auto revert patterns detected: 560
Actual reverts inside auto revert patterns detected (%): 204 (36.4%)
Total revert commits in period: 601

Revert categories:
  nosignal: 215 (35.8%)
  ghfirst: 151 (25.1%)
  uncategorized: 105 (17.5%)
  ignoredsignal: 70 (11.6%)
  weird: 46 (7.7%)
  landrace: 14 (2.3%)

Total reverts excluding ghfirst: 450
Reverts (excluding ghfirst) that dont match any auto revert pattern detected (%): (268) (59.6%)

*********************************************************************
STATS SUMMARY:
 PRECISION: 36.4%
 RECALL: 33.9%
 F1: 35.1%
*********************************************************************

Per workflow precision:
  Lint: 50 reverts out of 60 patterns (83.3%) [excluding ghfirst: 46 (76.7%)]
  trunk: 40 reverts out of 74 patterns (54.1%) [excluding ghfirst: 37 (50.0%)]
  pull: 79 reverts out of 276 patterns (28.6%) [excluding ghfirst: 74 (26.8%)]
  inductor: 34 reverts out of 144 patterns (23.6%) [excluding ghfirst: 31 (21.5%)]
  linux-binary-manywheel: 1 reverts out of 6 patterns (16.7%) [excluding ghfirst: 0 (0.0%)]
  ```

Author: jeanschmidt

aws/lambda/pytorch-auto-revert/pytorch_auto_revert/autorevert_checker.py

@@ -125,27 +125,38 @@ def _fetch_workflow_data(self):
             f"Fetching workflow data for {len(self.workflow_names)} workflows since {lookback_time.isoformat()}..."
         )
 
-        # For pattern detection we consider non-restarted main branch jobs only
-        base_where = "workflow_event != 'workflow_dispatch' AND head_branch = 'main'"
-
-        query = f"""
-        SELECT
-            workflow_name,
-            head_sha,
-            name,
-            conclusion,
-            status,
-            torchci_classification.rule AS classification_rule,
-            created_at AS workflow_created_at
-        FROM
-            workflow_job FINAL
-        WHERE
-            workflow_name IN {{workflow_names:Array(String)}}
-            AND {base_where}
-            AND created_at >= {{lookback_time:DateTime}}
-            AND dynamoKey LIKE 'pytorch/pytorch/%'
-        ORDER BY
-            workflow_name, workflow_created_at DESC, head_sha, name
+        query = """
+            SELECT
+                wf.workflow_name,
+                wf.head_sha,
+                wf.name,
+                wf.conclusion,
+                wf.status,
+                wf.torchci_classification.rule AS classification_rule,
+                wf.created_at AS workflow_created_at
+            FROM workflow_job AS wf FINAL
+            INNER JOIN (
+                -- Deduplicate pushes by head_sha using group+max,
+                -- keeping the most recent timestamp
+                -- this is faster than using distinct
+                SELECT
+                    head_commit.id as sha,
+                    max(head_commit.timestamp) as timestamp
+                FROM default.push
+                WHERE head_commit.timestamp >= {lookback_time:DateTime}
+                AND ref = 'refs/heads/main'
+                GROUP BY sha
+            ) AS push_dedup ON wf.head_sha = push_dedup.sha
+            WHERE
+                wf.workflow_name IN {workflow_names:Array(String)}
+                AND wf.workflow_event != 'workflow_dispatch'
+                AND wf.head_branch = 'main'
+                -- this timestamp should always be bigger than push_dedup.timestamp
+                -- it is just a optimization as this column is indexed
+                AND wf.created_at >= {lookback_time:DateTime}
+                AND wf.dynamoKey LIKE 'pytorch/pytorch/%'
+            ORDER BY
+                wf.workflow_name, push_dedup.timestamp DESC, wf.head_sha, wf.name
         """
 
         result = CHCliFactory().client.query(
@@ -207,14 +218,15 @@ def _fetch_commit_history(self):
         lookback_time = datetime.now() - timedelta(hours=self.lookback_hours)
 
         query = """
-        SELECT DISTINCT
-            head_commit.id as sha,
-            head_commit.message as message,
-            head_commit.timestamp as timestamp
-        FROM default.push
-        WHERE head_commit.timestamp >= {lookback_time:DateTime}
-          AND ref = 'refs/heads/main'
-        ORDER BY head_commit.timestamp DESC
+            SELECT
+                head_commit.id as sha,
+                head_commit.message as message,
+                max(head_commit.timestamp) as timestamp
+            FROM default.push
+            WHERE head_commit.timestamp >= {lookback_time:DateTime}
+            AND ref = 'refs/heads/main'
+            GROUP BY sha, message
+            ORDER BY timestamp DESC
         """
 
         result = CHCliFactory().client.query(