This import used to be on top of the file, but that creates a cycling
import.

preparing for the cython code, I used to grab the wrong inputs ( it should
be the node inputs not self.inputs)

reconstruct_graph always replaces CudaNdarrays with TensorTypes, so in case
of a gpu scan op, we can not use that logic to generate a hash ( I compute
it locally when I am generating the gpu scan op).

The way it works, __init__ gets a lambda function that construct a Tensor
Type. By default constructs Tensor Types, but move to gpu will replace it
with a CudaNdarray constructor.

Notes on this change: 1) I needed to replicate the inplace optimization, to
make a version that can handle ops moved on the gpu. I've also added a
replica of the reconstruct_graph function from scan_utils that again does
not replace CudaNdarray with TensorTypes ..

Two changes. First one is to make profiling work with scan (scan was looking
for a subclass of the ProfileStat object). Second, I made a more fixed
length printing of timings which I believe personally is much better.

While we agreed that there might be a more principial way of solving this,
this solution was fast to add and it is pretty efficient for now.

Is the shape of the data that should match.

In order for scan to run correctly inplace, it needs that none of the
initial states are the same memory buffer.

The old implementation used to result in stochastic order error in
debugmode. After many attempts to solve it, I decided that it would be
better and faster just to rewrite it. This new implementation does not
suffer from any bug (i.e. all tests pass in debug mode).

as Pascal suggested we should make it clear that scan doesn't have a normal
perform anymore, and that one needs to use make_thunk.

After talking to Pascal we decided that having this kind of funcion can be
quite useful, such that not every optimization does this splitting over and
over again ( it is really easy to have bugs by either messing up the order
or the count).

I think this commit got reverted ( I'm not sure when and why). Without it
graphs produced with pydotprint might be extremely misleading.

Small refactoring at the same time.

The formula before was wrong, it does not work when inputs are matrices,
because it depends on which axis y is computed.  I believe this formulation
is much better.

Before only the gradient of inc_subtensor was ok ( and it also was the only
one checked)

numpy.all doesn't check dtype and shape