Properly nest ThreadContextElement

[qwwdfsad]

* Restore the context in the reverse order of update, so they are properly nested into each other
* Also, do a minor cleanup

Fixes #2195

[qwwdfsad]

I've also measured the potential overhead and it turned out that even tho we allocate slightly more, the average throughput of context restoration (ChannelSinkBenchmark.channelPipelineTwoThreadLocals) has been improved by 20%!

For single-shot coroutines without hot loops, the performance is roughly the same (few percents better, actually) with the cost of an additional small allocation.

As an observation of the second order, it seems like the CoroutineContext.fold is not as fast as we tend to think

[elizarov]

Great improvement.