1) scipy is as fast as our implementation now (+- 10%)
2) the optimized graph hides shape mismatches (and currently can cause a
segfault)

refactored convolution test on the gpu to don't execution version of code that will execute the refenrece version because the optimization can't be applyed.

Made the type of exception raised in that case an InconsistencyError (was
ValueError) and made the GemmOptimizer catch that exception when trying to
insert gemms into a graph.

we do on purpose.

Also moved printing of the remainder of the message to INFO level logging, which
might be a bad idea.
But nvcc spews things that are not errors to stderr which are not easy to filter
out.  I'm thinking of the warnings that _POSIX_SOURCE is redefined, followed by
a multiline header traceback.
We do not want to print that to stderr by default every time someone compiles
anything with nvcc.

If compilation fails, then the whole nvcc stderr goes to our stderr anyway.

try and draw 6 samples per thread instead of one.