When an object is being fetched from the backend, other requests for it

When an object is being fetched from the backend, other requests for it
are parked not holding a worker thread, until the fetching session lowers
the busy bit on the object.

At that point we would release all the parked requests into the thread
pool to be serviced, causing instant congestion and calls for road-pricing
if the backend were a bit slow on an oft-requested object.

Change the restart policy to be paced exponential:

When we clear the busy bits, we unpark "rush_exponent" requests into
the thread pool to start the show.

Whenever the object is dereferenced, in practice whenever a request
has been serviced, another "rush_exponent" worth of requests are
unparked into the tread pool.

Set the parameter to a conservative 3 until we know more about the
behaviour in practice.

If it is a big object and/or the clients are on slow lines, 3 may
be an order of magnitude to small.

Attempts to fix: #188

git-svn-id: svn+ssh://projects.linpro.no/svn/varnish/trunk@2294 d4fa192b-c00b-0410-8231-f00ffab90ce4