Use RetainableByteBuffer in HttpConnection

[gregw]

The org.eclipse.jetty.server.HttpConnection class should use org.eclipse.jetty.io.RetainableByteBuffer rather than it's own reference counting of buffers. This will eventually allow a more efficient buffer pool abstraction to be used that provides RetainableByteBuffers directly.

[gregw]

@lorban Simone suggested that we do this. I had a quick look at the input side at least looks achievable, so I thought I'd punt to you with your HttpInput experience. I think it can be done entirely in HttpConnection since every increment of the reference counter is done by calling newContent, so the Content itself can do the retain and release. I've not yet looked at the output side.

[sbordet]

@gregw I don't think the output side is involved at all, as RetainableByteBuffer is only for the input side.
By using it, we will make all protocol implementations homogeneous and we could pool them more efficiently.

On the output side, all implementations currently rely on pooled "naked" ByteBuffers but I'm not sure how much better we can do considering that we have to wrap byte[] most of the times.

[gregw]

Once the next release is out, we need to ask for a few ByteBufferPool dumps from real deployments so we can plot histograms of buffer sizes in the pool.