False lock-order-inversion reports for locks taken in a single thread

Originally reported on Google Code with ID 81

```
Consider there're two locks, L1 and L2, and a thread T1 that does the following:
lock(L1)
lock(L2)
unlock(L1)
lock(L1)
unlock(L1)
unlock(L2)

, while other threads may (or may not) acquire and release L1 and L2 (but not both)
at the same time.
Based on the lock acquisition pattern in T1 ThreadSanitizer will report a lock-order-inversion
(potential deadlock), although there's no possibility of a deadlock as long as other
threads do not take both L1 and L2.
The error can even be reported in a single-threaded process, which isn't correct at
all.
```

Reported by `glider@chromium.org` on 2014-10-09 09:25:52