A benchmark object can specify a list of cmds to execute as
"servers" in the member attribute server_cmds.
Servers are started and terminated through Popen objects.
This requires the server cmds to not daemonize so the server
can be terminated through the Popen object.
For each started server cmd a shutdown function is registered with
atexit to terminate all servers even if a exception occurs.

Use the new server concept in httpd and mysql benchmarks

--analyze uses malt to trace the benchmarks behavior. It uses the run
loop but the obtained results are not stored

Benchmark.server_benchmark is used if only a server is started for
each allocator and clients are used to measure its performance in the run loop.
If server_benchmark is set to True the cmds are run with the system default
allocator.

Misc changes:

* The global environment is no longer changed. Instead a custom env dict
is passed to suprocesses containing LD_PRELOAD.

* Failing cmds no longer skip the whole benchmark instead they now skip the
malfunctioning allocator.

* Fix default title in plot_single_arg

an analyse run are not stored

bench now loads all *.py files from src/benchmarks as benchmarks