This annotation forbids starting states, where a non-preemptable subtask
is running but has not the highest priority in the system.

The TIMING_POINT_NO_INTERRUPTS_IN_BLOCK annotation forbids the action of
interrupts in the end or begin block to be included into the state
graph.

It seems that clang does not save r29 correctly, therefore we also save
and restore it. I'm unsure whether this is a bug somewhere in dOSEK,
somewhere in LLVM, or somewhere else where the light of the moon shines
upon us. Nevertheless, seems to work now.

The interrupt reaction benchmark is now also functionally executable.

The TIMING_POINT_START_INTERRUPT_IN_BLOCK marker results in a state
graph, that starts at all interrupt entries that originate from the
marked computation block.

a patch in the pasim repository (patmos/simulator) is required to use a
memory-mapped perfomance counter that returns the executed instructions

I should have committed earlier. But I didn't; I feel ashamed and i
admit, that it was a bad idea in the first place. So, my dear reader,
what did happen here:

- We detect loops on the State level and put that information into the
  PML. For this, an dependency to networkx was introduced, I should make
  this optional in the future.

- timing_start and timing_end can now take flags that influence the
  detection of OS-STG regions. Especially the TIMING_STOP_BEFORE flag is
  important for exit nodes. It has the semantic, that this ABB is NOT
  part of the circuit, but all states that would dispatch here are part
  of the circuit.

In order to provide an less intrusive (in terms of the WCET analysis),
the dumping of the timing circuits is activated and hidden into trap #2.

This time it also handles the copter benchmark.

How long does an task activation from an interrupt take?