Also fixed a bug where the same key.pkl file could be loaded multiple times.

We should use the linker=c|py_nogc flags.
Remove reference to FAST_RUN_NOGC that never existed.

PROFILE_MODE a suggestion message says you should compile the function in a
different mode like for e.g. FAST_RUN_NOGC. When you try to compile in this
mode, you get an error saying there is no such mode .. though there actually
is.

tensor.as_tensor_variable on inputs of x, but I should do that only if they
where numpy.ndarrays (maybe other types) .. because they could be random
states case in which tensor.as_tensor_variable fails.

instead of int .. not sure why but I don't see anything wrong with that

The idea is that even when the key changes, the resulting code and compiled
module can remain the same. We avoid this by keeping a hash of the C code (as
well as of the compilation options) and when this hash matches that of an
existing module, we re-use this module instead of having two different modules
doing the same thing.

Two improvements should be made to the current implementation:
    1. Currently a module is still compiled before we detect that it is
       duplicated, while we should avoid this compilation step.
    2. It may be the case that different code or compilation options actually
       lead to the exact same compiled module. It would be nice to be able to
       detect this situation as well.

found out this is not the right way of doing it, and there is a much
better(tested) way, namely specify_shape

of scan to be in a certain form, which is not true anymore if constant
folding replaces them by a constant.

node in a graph.

madness of elemwise operation introduced by infershape of subtensor and by
scan and makes plots of computational graphs readable and possibly faster.

check if they still use as little memory as needed in FAST_RUN.

gradients.

the key ( in this case the right key .. for scan to figure things out).