I used to call as_tensor_variable which is not ok for sparse or
random generators

node.

sequences are not a list of dictionary anymore, but a list of tensors,
the original tensors.

Now scan requires the user to say the number of steps. In case of outputs
that do not feed anything in, if the user does not provide any information
then it is assumed that you want all intermediate steps.

Is suppose to give a lot more control to the user, making it possible to
avoid some graphs that are hard to optimize, or which do not get optimized
properly. It is also suppose to generate much cleaner and small graphs.

Link blas to epd library correcly on windows.

Sparse

Remove use of sys.maxint in slices.

Sparse

vars with broadcastable dims were being disallowed unfairly from being update exprs

Added MulSD, AddSS optimizations and new blas function declerations to sparse sandbox

Remove the useless Avanced[Inc]Subtensor args and added missing fct.

Fix blas