added floatX doc.

now when we import theano.sandbox.cuda, we don't register gpu optimization in fast_run and inplace. When we call theano.sandbox.cuda, we add gpu optimizer to fast_run and inplace. This way, they don't get applied when use() is not call.

in class DB moved some stuff to a new fct add_tags(name,*tags). This allow to add tag to an optimizer after it have bean registered.

make local_gpu_sum raise an exception when the sum pattern is not implemented. It make the optimizer catch this error instead of failing at compile time.

added new optimizer InputToGpuOptimizer that transfert the input of a graph on the gpu card when needed. This make the gpu code work when not using shared variables

PermuteRowElements and ARange.

Added test to check for that (gradient included).
New definition of inverse_permutation, using PermuteRowElements and ARange Ops
instead of InversePermutation Op.

dtype.  This avoids annoying compiler warnings during tests.

This causes the conversion tests to pass (now that the code compiles) with
recent gcc (on sencha machine in the lab).

implementation.

Fix #236.

fix bug with local_gpu_shape. Their is GPUShape that inherit from Shape and it should return the good value.

cast to/from floatX that are not needed as they don't change the type will not generate an OP. Added test for this.

implemented the new flag THEANO_FLAGS=gpuelemwise.sync that default to True. When false it won't wait for the gpu fct to finish and won't check the return status. cuda/tests pass in debug mode. Should we add a sync at the end of all the theano graph? How to do so?