Write an optimizer that also checks correct usage of attributes

Tests for correct semantics:
- What are the inputs/outputs of the code?
- Are all parameters/initial values/differentials assigned a value?
- Are there any assignments to variables that have the wrong role for writing
  (illegal)?
- Are any variables assigned twice (illegal)?
- The methods of a flat object must not use any data from outside of the flat
  object.

Optimizations:
- Create missing assignments in 'if' statements
    - Makes creating all attributes at the beginnig of the function unnecessary
    - May remove problems with local variables
    - Maybe unnecessary when good dataflow analysis algoritm exists
- Remove variables:
    - All variables that are read only once
    - Variables that are never read.
    - Special attention needs to be payed to "if" statements.
- Remove statements that compute removed variables
- Reorder statements.
Algorithm could go backwards from the outputs.

Issues:
- Local variables need special design:
    - True for variables and automatically created variables (See Bug #600786,
      Bug #597234)
    - They should not always need a value (Goes this away when missing variables
      in if statements are generated?), currently problematic for arameters.
    - They also should not automatically be stored.
- Inheritance of role needs to be more refined.
    - RoleStateVariable is not inherited. It should become: RoleVariable
    - RoleParameter can only be inherited in initial functions.
      Assignments to parameters re illegal in "dynamic" and "final".
    - Alternative: special role for local variables:
      They can get any AttributeRole as long as they are not stored
      but potentially optimized away.

Libraries:
The following libraries could be useful to implement these features:

- python-graph: a pure python library with graph data structure, algorithms, and visualization via Graphviz.
  http://code.google.com/p/python-graph/

- Network X: a (pure python?) library with graph data structure, algorithms, and simple visualization. Maybe more complete than ``python-graph``.
  http://networkx.lanl.gov/

- Igraph: library with many algorithms. Not pure Python.
  http://igraph.sourceforge.net/documentation.html