Fix data loss on cancelled io

[mstetsyuk]

Example of a potential data loss:

1. We called FiberScheduler::read to read from the fd.
2. Then called future->cancel().
3. The cancel lands in the gap between the kernel writing the CQE to the ring and the scheduler processing the CQE.
4. The cancel does future->result = nullptr.
5. The scheduler does:

            int result = cqe->res;
            if (result >= 0)
            {
                if (future->result)
                {
                    *future->result = static_cast<uint64_t>(result);
                }
                future->set(0);
            }

6. Effectively, we read data from the fd, but "dropped it on the floor", and subsequent reads won't read it.

It's enough to just remove future->result = nullptr from FiberScheduler::cancelIo. And it doesn't seem like that will break any invariant.

To add a deterministic red-green test, the simplest solution I found is to inject a hook on a future which runs after reading CQE from the ring and before processing it. This code is only compiled for test builds (i.e. debug or sanitizer). Might be potentially useful for other similar tests as well.

[vadimskipin]

As I understand this is a cancel/completion race. Yes, we must not clear result pointer and let kernel to resolve the outcome.

[vadimskipin]

I am not an experienced github user - can you squash commits before merging?