During node crash, two concurrent modifications in can both succeed in pessimistic tx cache.
This also causes random failures in InfinispanNodeFailureTest:
1. TX1 originating on A acquires lock for key X, A is primary owner
2. C is killed and B becomes primary owner of key X
3. TX2 originating on B acquires lock for key X, B is now primary owner
4. TX1 commits the tx, Prepare is sent with the new topology id so it commits fine
5. TX2 also commits the transaction
Log attached (this is not master but changes should not be related).