Implement systematic tests for exception handling along specification.

This fixes redmin issues #8374, #8520 and #8364.

This was a left-over from times when no proper initial value propagation
and exception handling was implemented in ApplicationCore. Waiting until
the backends are open is anyway not sufficient, since the DeviceModule
needs to finish the recovery procedure. If any test would still rely on
this (anti-)feature, it would be subject to race conditions.

The wait is removed because it prevents important tests for the
exception handling specification. In particular, tests which
intentionally break devices during startup of the application would not
be implementable.

When writing special modules (e.g. the MicroDAQ module) it may be necessary to both override and call this function explicitly.

add Application::dumpModuleConnectionGraph()

This generates graphviz code to visualise the connections between modules

Otherwise people using this might be confused that their server starts
and does nothing

Require e.g. when needing to report an exception programmatically

The previous implementation using the mutex had two issues:

1. The mutex was locked in a different thread than unlocked (DeviceModule constructor vs. handleException() thread). This results in undefined behaviour.
2. The boost::shared_mutex::lock_shared() function is not an interruption point, so applications did not terminate while waiting for initial values. (boost::latch::wait() is an interruption point.)

Just use enableDebugDataLoss() to enable debug output.

This will print always a lot of information if one or more devices is unavailable at server start. There is a high risk that these messages distract from the actual issue (broken hardware), hence it is better to remove this.

The message of this exception is usually lost and the user is left with
an obscure error about not calling shutdown() in the constructor without
any clue what was wrong with the config

- Have to use float because it that's what is descibed in the map file.
  We are skipping the lmap which would adapt firmware types to the application type, so the type has to match
  The type has changed because it was always wrongly set to double for all floating types data, although the catalogue suggests float. This fix requires the test to be adapted.