perf: parallel io.Copy

[mmatczuk]

The io.Copy implementation is synchronous. It reads and writes in a for loop on a single thread. When read and write costs are high (i.e. using net.Conn) it cal halve the performance.

We could easily optimize that by prefetching data while writing.

Now

R R R
 W W W

Desired

RRR
 WWW

Implementation idea:

• Reader uses two buffers 128K each
• If buffer capacity goes below 32kiB the buffers are switched
• Reader sends buffer subslices to writer over a channel, channel capacity is 1
• Writer writes what it gets over the channel as in io.Copy

Note use subslice with capacity [start:len:cap] syntax.

[mmatczuk]

The idea was implemented in https://github.com/saucelabs/forwarder/pull/834 but did not yield the expected results.