The following changes cleanup the eval loop and result in a pretty solid 2 to 3% improvement in pybench for me.

And it is about 5% faster for long generators.

* Change why enum type to int and #defines.  And moved the why defines to opcode.h so that they can be seen by the genrator objects after a yield, return, or exception.

* Added an "int f_why" to frames so the generator can see why it returned from a send.

* Refactored generator obj so it can use the "f->f_why" to determine what to do without having to do several checks first.

* Moved the generator specific execption save/swap/and clear out of the cevel main loop.  No need to check for those on every function call.


The only test that fails is the frame size is test_sys.  I left that in for now so someone could check that, and tell me if it's ok to fix it, or if I need to do something else.

I also considered putting the why on the tstate object.  It might save some memory as there wouldn't be as many of those.