The select algorithm gets into live-lock in a corner case

[ndkoval]

In case we have two channels (c1 and c2) and two coroutines, which perform select on this channel but in the different direction, a livelock (StackOverflowError in practice) can occur.

Let's assume that the 1st coroutine processed c1 and then c2, while the second --- c2 and then c1. The following interleaving can occur.

1. Coroutine 1 goes to c1 and adds itself to the waiting queue;
2. Coroutine 2 goes to c2 and adds itself to the waiting queue;
3. Coroutine 1 goes to c2, decides to make a rendezvous, creates the corresponding descriptor and write it to its state;
4. Coroutine 2 does the same (wants to make a rendezvous on c1).

After that, both coroutines tries to select each other, for what each of them has to update the state of the opposite coroutine to SELECTED. However, they see descriptors, so they try to help each other, see descriptors again, try to help each other again, ... This way, they get into a livelock, what causes a StackOverflowError, because helping is written using recursion.

This could be easily fixed by having the total order on all SelectInstance-s, and writing a descriptor to the first in this order SelectInstance at first, and only then to the second one.