arcswap: lock vertex *before* gain calculations (and other things)

[hhirtz]

Otherwise there can be concurrent moves on neighbors V and U when:

1. Thread A computes the gain of V
2. Thread B computes the gain of U
3. Thread A locks V
4. Thread A checks for locks on neighbors of V -> there are none because B didn't lock U yet
5. Thread A moves V and releases the lock
6. Thread B locks U
7. Thread B checks for locks on neighbors of U -> there are none because A released the lock
8. Thread B moves U