faster retain/release in multithreaded apps, NSAffineTransform

[Richard Frith-Macdonald]

I've added code to NSObject.m in subversion trunk  to use atomic  
increment and decrement for the retain count rather than using  
locks.  this should make a big increase to performance in  
applications with a lot of active threads.
The new code should work on gnu/linux intel machines and on mswindows  
machines, but all other systems should fall back to using locks.
Could people with other systems please.
1. try it out to make sure that the fallback to use locks works  
correctly.
2. if possible, contribute code to use atomic increment and decrement  
instructions on their machine.
The use of the atomic operations should be pretty clear, just search  
for 'GSATOMIC' in NSObject.m
Of course, spotting any mistakes would be good too.

The other tricky thing today was porting NSAffineTransform  
optimisations from mySTEP ... the changes there will require a  
rebuild of the gui andfback libraries since the size of  
NSAffineTransform objects has changed.  Also, while I definitely  
fixed a couple of bugs (one in the original code and one in the  
mySTEP code), I could have introduced others ... please look out for  
any reproducible oddities.