Propagate errors from Accelerate back to Haskell

[tmcdonell]

Originally mentioned here:
https://groups.google.com/forum/#!topic/accelerate-haskell/UUNYssdEykQ

[amigalemming]

So far, I can at least simulate an error by an out-of-bound access to an array.

[yongqli]

Yes, this would be useful for me as well.

Also, a way to return NaN would be nice. Currently A.constant 0 / A.constant 0 causes nvcc exception: error: identifier "NaN" is undefined