Our transfer call are synchronized version. So no problem there. The
problem that we need to work around is that the Theano gc could free
the output var before we finish with it. cudaFree is instananeous, it
don't get in the stream of command to execute.

Documentation improvements

Consistent & correct handling of integers and gradients

-Documentation and implementation of a consistent way of handling gradients and integers
-Type checks that ensure the gradient is always floating point and not an integer
-Type checks that ensure the gradient of an integer is always undefined or 0
-An upgraded version of connection_pattern that provides theano with enough information to answer questions like "is variable x a function of variable y?" accurately

Correct Theanos's tutorial: add cross-references, correct some more typos, improve style a little, correct the logical structure some more

Correct Theano's tutorial: overhaul logic of exposition and integrate exercises and useful info found in CIFAR SC2011 accordingly

fixed an error message

added doc and argument checking to MRG_RandomStreams.multinomial