it was a mistake due to the list of list approach for the storage maps

The mintap is an offset that should always be taken into consideration

the to_list function used to be called only _list

I've cerated similarly to aux inpute a list of state_buffers, the difference
being that know we keep track of the length of each argument as well..

Even if we do not have access to the memory buffer, if we are dealing with
Tensor Types we can create a fake shared variable (all we need to know is
the number of dimensions and dtype). I've wrote a function that create such
a shared variable (that could afterwards take the real values) and used it
for the auxiliary inputs (paramters and inputs/sequences)

names were getting to long so I created some intermidiate lists of buffers

at make_thunk we don't actually have access to the memory buffer to be able
to do such things .. those have to be done at runtime

This op takes care of the special case when the entries of scan are not in
the righr order. Such a thing can happen once the save mem optimization is
implemented, and the memory buffer for an output is much smaller then the
number of steps. Then scan will just roll around that memory buffer, and
entry 0 might end up somewhere in the middle

Before this class was used to validate that computing the shape of an
argument does not rely on any input of the inner function of scan. We do not
need to do this anymore, since now the output storage size is given as input
to scan, and output shapes are the same as input shape (so the internal
graph of scan does not come into play anymore).

This way we make sure that no gpu argument gets there