fix: Serialize entry writes

[harrytmthy]

Summary

Fixes a rare but critical issue where overlapping .apply() and .commit() operations on the same SafeBox instance could cause data corruption or AEADBadTagException. The root cause was concurrent writes to the underlying blob store, especially when commit() arrived before a prior apply() had finished persisting to disk.

Implementation Details

This patch introduces serialized write sequencing via a per-instance writeBarrier: AtomicReference<CompletableDeferred<Unit>>. Each new write operation, from either .apply() or .commit(), will now:

• Wait for the previous write to complete (previousWriteBarrier.await()).
• Perform the write under writeMutex to ensure mutual exclusion.
• Complete its own barrier, unblocking the next writer in the queue.

This guarantees that apply and commit operations are serialized and isolated, preventing overlapping writes and eliminating the observed race conditions.

Test Adjustments

• Removed the legacy closeWhenIdle state test that asserted CLOSED emissions. On the 1.1.x line, close() and closeWhenIdle() are deprecated/no-op and will be fully cleaned up in 1.2.0 (see Enforce singleton SafeBox per file to improve fail-safety and prevent race conditions #58).

Notes: No public API changes, behavior is limited to internal write sequencing and stability.

Closes #60