use (idx,apply) for dictionary key in profile mode as (idx,op) cause conflict when their is many theano fct.

pydotprint now take an optional parameter mode. If a ProfileMode, print time stats in the apply on the graph.

added in the profiler a list of op that don't take float64 as input but generate float64 outputs. Usefull for the gpu.

enable cuda test when cuda is available. make some helper fct to use theano._asarray when needed and make the code more readable.

 Fixed a bug involving shape_i op and constant folding. In short, new constants introduced by the constant folding optimization were not inserted in shape_of dictionary

added device=gpu. Needed on angel as the gpu are in exclusive mode and we want the driver to select the gpu.

makes the directory containing libpython visible to the linker, just in case it doesn't get picked up otherwise. Not the most beautiful solution!

Two bugs fixed. First one was in printing, pmsg was spelled as psmg. Second one is more of a temporary fix. neg(x) was optimized to -1*x, where was dtype int which ended up in a dtype error in dot22scalar. I changed the optimization to create a -1 of same dtype as x but fancier solution possible.