I was processing the transitive inferred ranges in fold_stmt when it was 
the final statement in the block.  the substitute_and_fold engine 
actually does a bit of work before calling fold_stmt.  this patch moves 
the check to pre_fold_stmt instead so it gets done before the final 
statement in the block is processed... as was the original intention.

I also changed it so we always do this just before the last statement in 
any block.  This allows us to get transitive inferred ranges registered 
for returns, as well as just normal blocks which can feed other blocks. 
   Performance impact is minimal.

Bootstraped on x86_64-pc-linux-gnu with no regressions.  Pushed.

Andrei