persistit: keep JournalManager consistent when rollover fails near the block end

[vharseko]

Summary

Fixes #265.

Under sustained Thread.interrupt() stress the JournalManager write-buffer bookkeeping could be torn around a journal rollover, after which every journal writer tripped the invariant assert (writeBufferAddress=2,999,999,998 position=0 currentAddress=3,000,000,038). The failure is two-phase:

1. Poisoning. rollover() is not exception-safe: writeJournalEnd() advances _currentAddress by JE.OVERHEAD (40 bytes) to within 40 bytes of the block end, and only then does the I/O (flush/truncate/force). An InterruptedIOException there (surfaced by MediatedFileChannel on an interrupted thread) abandons the rollover before _currentAddress jumps to the next block boundary. The invariant still holds at this point, but the reserve prepareWriteBuffer maintains for the JE record is gone: fewer than JE.OVERHEAD bytes remain to the block end, and the write-buffer limit is clamped to that remainder.
2. Tearing. The next journal write retries the rollover. The JE.put* helpers write into the backing array directly, bypassing the ByteBuffer limit, so nothing fails until advance(JE.OVERHEAD) updates _currentAddress first and then throws from _writeBuffer.position(...) (newPosition > limit), leaving _currentAddress exactly 40 bytes ahead of _writeBufferAddress + position - the observed values. Every subsequent journal write then fails the assert (or hits the same IllegalArgumentException with assertions disabled).

Changes

• JournalManager.writeJournalEnd(): skip the JE record when it no longer fits before the block end, so a retried rollover completes and heals the poisoned state; recovery treats the missing JE as a dirty shutdown.
• JournalManager.advance(): update the buffer position (which validates against the limit and may throw) before _currentAddress, so a failure cannot tear the invariant.
• JournalRecord: fail fast (IndexOutOfBoundsException) when a record write would exceed the buffer limit instead of silently writing past it into the backing array.
• IOFailureTest#testRolloverRecoversAfterFailureNearBlockEnd: deterministic reproduction - positions _currentAddress exactly JE.OVERHEAD + 2 bytes before the block end, aborts the rollover with an injected truncate() failure, verifies the poisoned geometry, then verifies the retried rollover heals and journal writes keep working.
• TransactionTest2: workers catch Throwable so an internal Error is reported and counted instead of silently killing the thread (the cascade that polluted transactionsConcurrentWithPersistitClose in the same CI run).

Testing

• The new test fails on the unfixed code with IllegalArgumentException: newPosition > limit: (40 > 2) (and the same error from tearDown, showing the state never heals) and passes with the fix.
• Full persistit/core suite: 566 tests, 0 failures, 0 errors, 5 pre-existing skips.