better error message

Fix bug with mixed types in shape of random sample

This fixes a bug when the shape is a tuple containing both symbolic
and numeric scalars.

Better error handling and buildbot

Do not return None in scalar op's grad, add tests

Fix cycle

Add case where 'axis' is a numpy scalar array

The python error is already set by device_malloc and it contain more information.

If the 'axis' variable is a numpy scalar array (like numpy.array(1)), these conditional branches will fail as it tries to build a set over the non-iterable array. I've added another branch to handle this case.

Also change checks to verify the output dtype of the Op itself, not of
the inputs or gradient, because it can depend on different things.
The idea is that if the Op's output is continuous, then the gradient
should be propagated to the inputs, regardless of whether they are
continuous or discrete. However, if the output is discrete, then the
gradient wrt the inputs will be a continuous zero.

instead of AssertionError, so it gets correctly detected by new test.

Mixed small stuff

If it happen frequently on the same test, we could modify this test to be more verbose or faster.