Test for broken copy-more-arg harder

Test for broken copy-more-arg harder

* It takes 17 fixed arguments for (threaded) PPC to fail.  Test
  for up to 33 arguments now.

* Also, add some comments to explain why SP shouldn't be set to its
  final value eagerly when the stack frame is smaller than the fixed
  arguments: accessing slots past SP is a bad idea when interrupts
  could hit and overwrite these values.  In such cases, leave SP
  pointing to the end of the source vector, and only move it back to
  the end of the destination vector after the copy loop.