Improve memory usage

[GoogleCodeExporter]

We should improve memory usage for large projects. Recently, Lucene's build 
crashed with OOM because too many classes were loaded. This was mainly caused 
by the task being run top-level, but forbidden should still not hold everything 
in memory.

The following improvements are easy:
- For related classes, we don't need to hold the ClassReader in memory (holding 
the byte array). Those classes are only parsed and signature lists are created. 
We should only extract superclasses and interfaces and throw the classreader 
away
- We may add lazy loading on classes to scan. This could be done by a type of 
"ClassLoader" that does a lookup on the Ant/Maven FileSet. The class is only 
parsed one time and classreader is thrown away. After parsing the class we just 
keep the sigantures for later lookups of related classes available.

Original issue reported on code.google.com by uwe.h.schindler on 10 Nov 2013 at 11:07