This adds a docstring for raise_with_op and also modifies its behaviour
somewhat. Previously, the exception was re-raised with a different
'args' member -- namely, in addition to the exception message it added
two new elements to the tuple: the Op itself and a string telling you
where in the graph the corresponding Apply node appeared.

This gets in the way if you want to have an exception message containing
formatting (newlines and tabs), since the entire tuple is printed, and
the formatting on the string is not rendered.

I've never seen the Apply node index be of use to anyone debugging, but
my proposed solution is to stash that information in __op_instance__
and __applynode_index__ on the object instead of in the args tuple.
I would not be against appending to args[0], the exception message,
with applynode information, but this is easy enough to obtain at the
pdb prompt with my modifications.

env.

This used to make pickling of scan impossible. As I've broken the pickling
mechanism several times before, I've added a test that checks if scan is
still picklable. I'm not sure is the best way of doing it though ..

review.

As before, I would suggest reviewing just the cython code and not the C code
since is generated, and hence hard to impossible to review.

while)

Cython can not handle negative indices ( because of a derictive at the
begining of the function that says not to do boundschecks). This is used to
make the code faster, so this workaround is better then removing the
directive.

scan)

All these changes have been generated by cython, and as such I don't think
they need to be reviewed but rather reviewing the cython code should be more
then sufficient.

In fact, it could hid problem that where not expected.