Fix following Pascal suggestion of the __eq__ of scan.

It helps by compiling the inner graph only once, making compilation much
faster

Basically change the code to be aware that the last input can be a
condition, and if so stop when that condition becomes True.

Specifically for initial states, they should always stay variables, and not
be constants !

My optimization pushes in the main graph a lot of computations in terms of
constants, and constant folding maxes out pretty quickly. This is a quick
fix to that. I'm not sure if 50 is not too much, I haven't played with that
value .. maybe 20 or something would suffice. I would also like to increase
this value only for constant folding and not all the others, but I didn't
know how.

theano.until

(note : the implementation of make_thunk is seen by DebugMode as a python
implementation, and it sees no C implementation for that Op. Are there
checks for the C implementation that are not done for the python
implementaion ? )

treat them like any other theano tensor)

Note: this optimization expects the state of scan to have the as_while flag
and the profile flag.

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