[#151] Fix the batch test race and return pooled connectors in UseCase3 tests

[vharseko]

Fixes #151

Problem

LocalConnectorInfoManagerTests.testBatchUseCase3Failure checked that there were no results right after executeBatch(). The UseCase3 test connector delivers results to the observer from its own thread, so the first onNext() can run before that check. Stress-running the test 300 times on master gave 6 failures with expected [0] but found [1], the error reported in #151.

Fixing it turned up two more problems:

• Pooled connectors were never returned. Each Subscription that executeBatch()/queryBatch() return holds a pooled connector until it is closed. The tests never closed theirs. A single run hides this, but repeated runs exhaust the pool and hang in ObjectPool.borrowObject().
• NPE on close. Once the batch has finished, the connector may return null from queryBatch(); TstConnector does. SubscriptionImpl wraps that null. Its getReturnValue() already handled null, but close() and isUnsubscribed() threw an NPE.

Change

• LocalConnectorInfoManagerTests:
  • testBatchUseCase3Failure no longer checks the transient state. It waits up to 3 s for hasError instead of a fixed Thread.sleep(500), the same approach Stabilize release-maven by removing race-prone assertion in testBatchUseCase3 #93 took for testBatchUseCase3. hasError is set after the result is added to the list, so both results are in place when the wait ends.
  • Both UseCase3 tests close their subscriptions in finally. testBatchUseCase3 keeps the queryBatch() subscription in its own variable instead of overwriting the executeBatch() one.
• SubscriptionImpl:
  • In the executeBatch() and queryBatch() wrappers, close() and isUnsubscribed() now accept a null subscription. So does getReturnValue() of executeBatch().
  • doRelease() still runs first, so the reference counter is always released.

Contrary to what #151 says, testBatchUseCase2 and testBatchUseCase2Failure are not affected. In UseCase2 the results reach the observer only through queryBatch(), which the test thread itself calls.

Testing

• Stress run, 300 invocations each of testBatchUseCase3 and testBatchUseCase3Failure: 600/600 pass in 65 s, with no hang.
  • Before this change: the old assertions failed 6/300.
  • With the leak still present, the run hung after about 10 invocations.
• Full test runs pass for connector-framework (197), connector-test-common (5), connector-framework-internal (509, 2 skipped as before), connector-framework-rpc (21) and connector-framework-server (32).