set_subtensor.

This was the cause of the 3 failing test in scan. I also revised the entire
optimization, writing a bit more comments/

variables and some other cosmetic changes.

* * *
More informative error message

ensure they don't belong to a different env.

Also update the documentation.

The scope is to make the computational graph more readable, and to move
constants in the inner graph, allowing constant folding in the inner graph
to do its job (performance).

obtain some outputs/or other node.

In order for scan to identify which of the inner outputs is the stopping
condition I use this dummy class to wrap my condition in it.

Look for conditions that the user might provide ( if scan is to terminate on
a condition ). This feature will be later used the conditional terminating
scan.

One thing that I've tried is to avoid importing cuda.

The main thing I was looking for was efficiency. The new algorithm
is not recursive, it considers a bunch of outputs at the same time,
and it uses toposort to traverse the graph, making sure each node is seen
only once.

safe_to_cpu was one of the reason that I was importing cuda. It was a
misunderstanding on my part, and the way is done now is somewhat more sane.