Only close AWS discovery clients that were actually built

[pjfanning]

Motivation

Follow-up to #907.

The ec2ClientUsed / ecsClientUsed flags added in #907 are set at the top of the lazy client initialiser, before the client exists:

private lazy val ec2Client: AmazonEC2 = {
  ec2ClientUsed = true          // <-- before the client is built
  ...
  builder.build()
}

A Scala lazy val whose initialiser throws stays uninitialised and is re-evaluated on the next access. So if builder.build() fails, or the custom client-config FQCN cannot be instantiated (throw new Exception(s"Could not create instance of '$fqcn'", ex)), we end up with ec2ClientUsed == true and no client. The PhaseServiceUnbind task then calls ec2Client.shutdown(), re-enters the initialiser and throws again — which is precisely the shutdown-time client creation the flag was added to avoid (see @He-Pin's review comment on #907). Setting the flag before builder.build() returns also leaves a window where shutdown observes true while another thread is still constructing the client.

Separately, the two collection changes in #907 are not the optimisations the description claimed:

• taskArns.grouped(100).toSeq — IterableOnceOps.toSeq is immutable.Seq.from(this), i.e. strict, and List is the default immutable.Seq. This materialises exactly like the .toList it replaced.
• EC2 getInstances — java.util.List.asScala yields a strict mutable.Buffer, not a lazy view, so flatMap then map build two intermediate buffers before .toList. No fewer allocations than the original.

Both paginated loops also do accumulator ++ page per page (List in EC2, immutable.Seq in ECS), which is O(pages²) copying.

Modification

• Replace both @volatile var flags with an AtomicReference that is populated after builder.build() succeeds. The shutdown task closes whatever client it finds there and returns Done when it is null, so it never touches the lazy val.
• EC2 getInstances: use .view so flatMap/map are genuinely lazy, and accumulate into a Vector.
• ECS describeTasks: flatMap straight off the grouped iterator instead of materialising the groups first.
• ECS listTaskArns: accumulate into a Vector so appends are amortised constant time.
• New Ec2ClientShutdownSpec and EcsResolveTasksSpec. EcsServiceDiscovery.resolveTasks is relaxed from private to private[ecs] so the latter can drive it — a new method in bytecode, so MiMa-clean.

Result

Shutdown never creates, or re-attempts to create, an AWS client — including after a failed creation. Pagination is linear in the number of pages rather than quadratic, and the EC2 chain allocates one list instead of three.

Tests

• sbt "discovery-aws-api/test" — 8 succeeded, 0 failed.
• Directional check: sbt "discovery-aws-api/testOnly *Ec2ClientShutdownSpec" with Ec2TagBasedServiceDiscovery.scala reverted to main — "should not re-attempt EC2 client creation during shutdown when creation previously failed" FAILED (2 configuration attempts instead of 1). Passes with the fix.
• sbt "+discovery-aws-api/mimaReportBinaryIssues" — success.
• sbt "discovery-aws-api/scalafmt" "discovery-aws-api/Test/scalafmt" and sbt headerCreateAll — clean. sortImports is not a task in this build (only javafmtSortImports); no Java sources changed.
• Not exercised against live AWS — no account available. The specs stub AbstractAmazonECS and ClientConfiguration so no credentials, region or network access are needed.

References

Refs #907