crossentropy_softmax_1hot_with_bias_dx

These two functions are used by as_tensor_variable to determine how to turn
python ints and floats into ndarrays for TensorConstants.

This provides an i-hope-not-too-hacky way for config.floatX=='float32' to make it so
that python literals like 1.1 don't force an upcast in expressions like
(fvector() + 1.1).

Another option would have been to leave the downcast of 1.1 in the graph as a
symbolic node that would be pre-computed at compile time, but I think that would
behave pretty similarly, and further burden the optimizer.

the optimization they register.  This way they can be used as decorators
without accidentally making their optimization disappear from the defining
module.

cross-entropy

added an not used argumement to DebugMode.__init__ to have it accept the same arguments as other Mode. This allow to work in more generic way.