doing the right thing during fini.

Patch by: Martin

Patch from Martin, with additional attention to VGC layout by me.

Available operations on processes are:
- start
- wait
- kill (send a signal)
- stop (kill with TERM)
- write (to its stdin)
- close (its stdin)

A process name starts with a 'p' and accepts a shell pipeline as the
command line. Environment variables can be set directly in the command
line.

A process will create a macro with its PID, and its stdout and stderr
are redirected to files in a dedicated directory.

Also add a varnish "name" macro in varnishtest.

(BO, REQ & VCA)

Add new a new TASK_QUEUE level for backend requests and give it
higher priority than client requests to avoid a priority inversion
when (almost) all threads are stuck waiting for backend fetches
which doesn't happen because there are no worker threads.

Backend polls are left at TASK_QUEUE_REQ priority for now, pending
reasons not to.

tasks from the queue limits.

Slightly change the queuelimits to be relative to the full complement
of worker threads in the pool, rather than the count of currently created
threads, this should drop less traffic during ramp-up, but retain the
steady-state behaviour.

an array of queues.

pools that service them.

Fixes #1742

It didn't look like the other options in both man pages and sphinx docs.
However, options like "-t <...>" are still somewhat broken in the sphinx
documentation: the option and description are on different rows.

Follows from the discussion of 49712a5f on varnish-dev@.

The development package contains vmodtool.py, which works better
if the interpreter is available.

Fixes:	#1746

It used to be possible to specify -sfile size in percentage. This
feature has been removed. Update the users guide to reflect this.

Fixes: #1744

Fixes #1745

shutuppery about indentation.

order.

Use it to health poke backends, rather than wasting a thread per
backend+poll.

for a very long time and never been shown to actually improve performance.

These days we pick busyobj out of a memory pool which gives us much better
control of the amount of memory sitting idle waiting to become a busyobj.

In case a single mempool not scale well enough, we make a mempool per
thread pool to scale horizontally.

config we have in Jenkins.

See comment XXX questions to discuss with phk