Buffer the output of gzip.Writer to avoid stalling

[jonjohnsonjr]

Use a bufio.Writer to buffer gzipped output while we are reading from
the other end of an io.Pipe to allow gzip to keep compressing its input.

A 64K buffer was chosen for its humor value. The default size of
bufio.Writer was too small when testing against a local registry.
Increasing beyond 64K didn't seem to have any noticeable effect. It
might make sense to make this smaller, but I don't see a reason to worry
about it ATM.

Fixes #920

In local testing, this improved write speed from ~8MB/s to ~30MB/s. Against an actual registry (Docker Hub or GCR), this seems to improve things by ~30%.

[jonjohnsonjr]

Thinking about it a little more, 64K is probably pretty good, actually. The Read()s seem to be happening in ~32K chunks, so having that plus some additional buffer should remove gzip as the bottleneck for a while.

[imjasonh]

💯

[jonjohnsonjr]

We might want to look at using a sync.Pool if we find that this causes the garbage collector to churn too much: https://github.com/containerd/containerd/blob/4d7d63f390e73f62676cd4994233805a6f821ab7/archive/compression/compression.go#L58

I was poking around containerd and docker to see if they do anything like this, in the hopes of landing a fix against them too, and it seems they do even better :P