[issue559] Solve the component interaction problem.

[FlorianPommerening]

This is not quite ready for review yet.

Notable changes:

• All components now exist in two varieties: task-independent and task-specific.
• For most components the task-independent variety is an instance of a template class that takes and stores constructor arguments, and then passes them on to the task-specific variant when needed (see component.h). For components where the general template doesn't fit, we add new classes for the task-independent version. Many of the classes we already have already use a factory pattern (landmark factory, pattern generator, etc.) - we plan to use them directly as the task-independent version to reduce the nesting level. But this is deferred to follow-up issues.
• Heuristic no longer takes a transform argument on the command line. Instead, there is a new heuristic cost_adapted_heuristic that evaluates a nested heuristic under different costs (so far this was the only publicly available transformation).
• Instead, all task-specific classes now take a shared_ptr<AbstractTask> and store a TaskProxy internally. This task is passed to their constructors when binding a task to a task-independent object. This is currently only used by heuristics and search algorithms. All other classes still get a task in the same way they did before (e.g., in generate_patterns(task)). This is because of the factory-factory pattern mentioned above and will be clearer once we get rid of it.
• We simplified the inheritance of Feature classes in the plugin code and use a TaskIndependentFeature where we previously used TypedFeature.
• Features now define options for the task-independent classes, so e.g. take an TaskIndependentEvaluator where they previously took an Evaluator. Our templated automatic task-independent class recursively binds those to a given task before passing them on to the constructor, so the task-specific classes still take the original arguments (e.g., Evaluator). The automatic documentation was adapted so the output should stay the same (still calling the argument and Evaluator).
• We got rid of a lot of parser code for copying ASTs and lazy instantiation which was only required for iterated search