Improve unbounded channels

[mahkoh]

I've added benchmarks and it turns out that comm channels are sometimes faster than stdlib channels. Unfortunately I messed up the commit message because I tried to use markdown syntax. Here is what it's supposed to look like:

# Where comm is faster than stdlib:

mpsc unbounded async_comm        45177
mpsc unbounded async_stdlib      51012

spsc bounded   async_comm        16160
spsc bounded   async_stdlib      19679

spsc bounded   sync_large_comm   31155
spsc bounded   sync_large_stdlib 67057

spsc bounded   sync_mid_comm     52558
spsc bounded   sync_mid_stdlib   68266

spsc bounded   sync_small_comm   424309
spsc bounded   sync_small_stdlib 544187

spsc one_space comm              7077
spsc one_space stdlib            8626

spsc unbounded async_comm        26718
spsc unbounded async_stdlib      30650

# Where comm is slower than stdlib

spsc unbounded sync_comm         36019
spsc unbounded sync_stdlib       29139

mpsc unbounded sync_comm         60856
mpsc unbounded sync_stdlib       44723

# Possible reasons

- Something in the sleep/wakeup handling: Maybe. But maybe not because
  our sync stuff is faster in the bounded channels.
- Node caching: The stdlib reuses old nodes instead of allocating new
  ones.

This is already looking quite good except for the unbounded+sync versions. Note that this doesn't touch Select and that the mpsc tests only test with two producers.

[marcusklaas]

Cool! Maybe you should try to merge with the standard library? There seem to be mostly advantages over the current implementation.

[mahkoh]

I don't believe channels should be in the stdlib.