Make sure each invocation of block_on uses its own Parker

[ghost]

This fixes a bug where we have recursive (nested) invocations of block_on:

task::block_on(async {
    task::block_on(async {
        // ...
    })
})

Each block_on behaves as an independent task and has an associated waker.

The problem was that both of these block_on invocations reused the same Thread descriptor for wakeups. If someone called .wake(), we had no way of distinguishing which of the two block_on tasks was actually woken.

Sometimes the outer task should get woken but the inner task would pick up the wakeup because it took over the Thread descriptor.

The problem is solved by associating a distinct Parker with each block_on invocation so that they don't get shared among recursive calls. Note that a Parker is just an implementation of std::thread::park() and std::thread::unpark(), except it was extracted out and moved into crossbeam-utils.

This should fix deadlocks @svartalf and @jebrosen have reported this week. Can you perhaps confirm the problem goes away with this branch?

[jebrosen]

Yes, this branch does appear to fix the deadlock I observed.

[svartalf]

Yes, I can confirm that it resolves the issue I had. Great work, as usual!