[1.8.x] Fix reused fork-join task submissions

[He-Pin]

Motivation

Backport #3621 to 1.8.x. Repeatedly submitting the same mailbox as a ForkJoinTask can corrupt a fork-join work queue through task identity reuse. Exceptional completion can also prevent subsequent submissions of that mailbox from running.

Modification

• Always wrap execute(Runnable) submissions in a fresh PekkoForkJoinTask in both Pekko's Java 8 and Java 9+ fork-join executors.
• Retain Mailbox extends ForkJoinTask and its existing methods for 1.x binary compatibility.
• Adapt the resubmission regression tests from fix: stop handing a reused ForkJoinTask to the fork-join dispatcher #3621 to cover both executor implementations, including recovery after an exception and a plain Runnable control. Wire TestJdk9 to the actor module's CompileJdk9 output.
• Document the fix and its built-in executor scope.

Result

Every submission through Pekko's built-in fork-join executors has an independent task identity and completion state. The mailbox binary shape is unchanged and no MiMa filters are needed. Custom executors that submit mailboxes directly to a raw JDK ForkJoinPool remain outside this backport's scope.

Tests

• Both regression suites demonstrate the failure on the original executor code: each reports one expected failure and one passing control. Both pass with the fix.
• Cross-Scala regression tests: 12 passed across Scala 2.12.21, 2.13.18, and 3.3.7, covering Test and TestJdk9.
• Runtime checks: 2 tests passed on JDK 8; both executor suites passed on JDK 25 (4 tests, compiled with JDK 17).
• Dispatcher, Mailbox, ForkJoin, and ActorSystem focused tests on JDK 17: 190 passed, 2 existing pending tests.
• +mimaReportBinaryIssues, +headerCheckAll, headerCreateAll, scoped sortImports, native scalafmt for changed files, and git diff --check passed.
• docs/paradox passed with -J--add-opens=java.base/java.lang=ALL-UNNAMED after the initial JDK 17 reflective-access failure. Initial disk-space and cross-JDK launcher/compiler setup failures were resolved before the successful checks above.
• checkCodeStyle fails on pre-existing formatting in remote/src/test/scala/org/apache/pekko/remote/serialization/NestedPayloadDepthSpec.scala; that file is identical to origin/1.8.x.
• PR_TARGET_BRANCH=origin/1.8.x validatePullRequest completed but was not fully green. LevelDB JNI lacks support for the Apple Silicon host, a Netty test cannot bind 127.0.1.1, and the aggregate run reported two TcpConnectionSpec timeouts and one FlowFlatMapConcatParallelismSpec timeout.
• Both TCP cases passed individually on the original 1.8.x code and on the fixed code. The FlatMap suite passed all 387 tests when rerun on JDK 17 with sbt's exported test classpath and the same time factor. These isolated reruns do not replace the failed aggregate result; no assertions or timeouts were changed.
• Auditable Qoder stdout review: No must-fix findings.

References

Refs #3621, #3620.

[He-Pin]

I think it's better just do it as 2.0.0