[ML] Sync changelogs between 8.4 and main (#89377)

Copies a few changelog edits from the 8.4 branch into main so that
people looking at the main branch see the same docs as people
looking at 8.4.

Relates #89078
Relates #89286
Relates #89376

Author: droberts195

docs/changelog/87361.yaml

@@ -3,4 +3,17 @@ summary: "Implement per-transform num_failure_retries setting"
 area: Transform
 type: enhancement
 issues: []
+highlight:
+  title: Infinite and adaptive retries for transforms
+  body: |-
+   Infinite and adaptive retries – available in 8.4 – makes it possible for 
+   transforms to recover after a failure without any user intervention. Retries 
+   can be configured per transform. The transform retries become less frequent 
+   progressively. The interval between retries doubles after reaching a one-hour 
+   threshold. This is because the possibility that retries solve the problem is 
+   less likely after each failed retry.
+   
+   In the *Transforms* page in *{stack-manage-app}* in {kib}, the number of retries 
+   can be configured when creating a new transform or editing an existing one.
+  notable: true
 

docs/changelog/88589.yaml

@@ -0,0 +1,21 @@
+pr: 88589
+summary: Make composite aggs in datafeeds Generally Available
+area: Machine Learning
+type: feature
+issues: []
+highlight:
+  title: Composite aggregations in datafeeds are Generally Available
+  body: |-
+    The support for 
+    {ml-docs}/ml-configuring-aggregation.html#aggs-using-composite[composite aggregations]
+    in datafeeds is now generally available.
+
+    [discrete]
+    [[early-stopping-dfa]]
+    === Optimizing speed of {dfanalytics}
+    {dfanalytics-cap} is even faster in 8.4. The new function automatically 
+    stops the process of hyperparameter optimization early in case of the 
+    accuracy gain for a different set of hyperparameter values would be 
+    insignificant. The early stopping of the optimization process results in a 
+    shorter runtime for the {dfanalytics-job}.
+  notable: true