Now the core of the execution model is committed, I've applied this patch to 
enable the nvptx dimension checking.  We force the vector size to be 32 in all 
cases, and check the worker size is not above 32.  Warnings are given if the 
user specifies an unacceptable dimension.

This patch  exposed some unacceptable testcases which I've modified to specify 
an acceptable vector length.   I also took the opportunity to fix up their 
reduction variable copying, which is only working right now because we default 
to copy, rather than firstprvate (that's the next patch for review).

Also a new testcase to exercise the error handling.

Committed to trunk.

nathan