Move multiplexer::elem to the request

[mzimbres]

I think there is no reason to keep multiplexer::elem in the multiplexer instead of the request. Moving it would also eliminate one dynamic allocation from each async_exec helping to reduce latency even further.

[anarthal]

There is, see #310. A cancelled request is implemented there as an elem without an attached request.

[mzimbres]

I have to look in more detail but I guess we could avoid that dynamic allocation somehow.

[anarthal]

I'm open to different implementations though. I think ee could replace the deque and shared_ptrs by linked lists, possibly using Boost.Intrusive. It would save the overhead of atomic counters, and the repeated partitioning in the multiplexer (which causes allocations too). But it means introducing a dependency which can't be hidden in .ipps.