Restore thread context elements when directly resuming to parent

[elizarov]

This fix solves the problem of restoring thread-context when returning to another context in undispatched way. It impacts suspend/resume performance of coroutines that use ThreadContextElement since we have to walk up the coroutine completion stack in search for
parent UndispatchedCoroutine. However, there is a fast-path to ensure that there is no performance impact in cases when ThreadContextElement is not used by a coroutine.

Fixes #985

[elizarov]

Marking it as a DRAFT. It works, but we need to evaluate its performance impact.

[qwwdfsad]

Performance impact on our stress for context and channels

Benchmark                                     Score change                                      
ChannelSinkBenchmark.channelPipeline                  -1%
ChannelSinkBenchmark.channelPipelineOneThreadLocal   -11%
ChannelSinkBenchmark.channelPipelineTwoThreadLocals   -7%

[qwwdfsad]

More benchmarks

Baseline:

Benchmark                            Mode  Cnt  Score   Error  Units
ChannelSinkDepthBenchmark.depth1     avgt   10  3.025 ± 0.085  ms/op
ChannelSinkDepthBenchmark.depth10    avgt   10  2.987 ± 0.082  ms/op
ChannelSinkDepthBenchmark.depth100   avgt   10  2.964 ± 0.041  ms/op
ChannelSinkDepthBenchmark.depth1000  avgt   10  3.012 ± 0.067  ms/op

Original change:

Benchmark                            Mode  Cnt  Score   Error  Units
ChannelSinkDepthBenchmark.depth1     avgt   10   3.400 ± 0.069  ms/op
ChannelSinkDepthBenchmark.depth10    avgt   10   3.793 ± 0.074  ms/op
ChannelSinkDepthBenchmark.depth100   avgt   10   5.147 ± 0.462  ms/op
ChannelSinkDepthBenchmark.depth1000  avgt   10  25.376 ± 0.851  ms/op

Recent change:

Benchmark                            Mode  Cnt  Score   Error  Units
ChannelSinkDepthBenchmark.depth1     avgt   10  3.165 ± 0.124  ms/op
ChannelSinkDepthBenchmark.depth10    avgt   10  3.141 ± 0.060  ms/op
ChannelSinkDepthBenchmark.depth100   avgt   10  3.178 ± 0.071  ms/op
ChannelSinkDepthBenchmark.depth1000  avgt   10  3.214 ± 0.105  ms/op