Add better documentation.

The problem does not get picked up in the unit tests, and only occurs with
a real dataset to work on. The problem is that we get a overloaded stack
due to some clusters becoming their own children. When the getLabel is
then called on the compositecluster, it results in an overloaded stack. I
think this could be fixed by changing the logic and working of the merge
method to use a member vector<set<label_id> > to track each cluster group
for the incomplete clusters, and only generating the composites (which
would only ever by one level deep) at the end of the process.

A composite cluster will fix the problem I've been having. Namely that the clusters are processed in parallel to get the uniform min, but the order that they are processed in will affect the end labeling. If clusters are set up as composites before hand. This will not be a problem.

A very simple solution to a threading issue. We can safely process the
peaks list in parallel if the integration can be guaranteed to be thread
safe. By making the interation a const member that simply returns the
results we do this. Therefore concurrent thread access will cause no side
effects. This Will allow us to handle the case where the user has selected
a threshold such that two peaks end up sitting on the same cluster and
therefore the integration method could be entered simultaneously.
Functional programming had the solution.

Also, I've made some changes to reduce the amount of work happening to
determine the Clusters which are unresolved. I can avoid this duplication
because for each indexpair that gets registered, there will always be a
duplicate (there has to be as each unreachable neighbour for one process
index is itself an unreachable neighbour for another process index) and
therfore only one cluster needs to be entered into the vector.

Have the integration working parallel and non parallel. Peformance tests
have reduced in speed by 1/2 so now run in approximately 1/2 a second. The
non-parallel version runs quicker than the parallel version indicating
that some work still needs to be done to improve the speed of the merging.

Cluster extracted to it's own type in it's own physical structure.
Clusters are now passed around. Tests added for the cluster type.