try and draw 6 samples per thread instead of one.

help it go faster though.

It was putting into the same profile the time taked inside the scan op to the sacn op AND into the individual op.

Now make multiple profiler instance to split them.

make shared constructor for scalar respect floatX and make test that tensor shared constructor respects it.

This is done to don't insert Rerebroadcast op that do nothing after those optimization are applied. This happen with GpuJoin.

make the api of the cuda shared variable constructor similar to the tensor shared variable constructor.

Added checks around call to CopyFromCudaNdarray in setitem, and added Equal method for comparing CudaNdarrays in C code

Added a few things to support GpuJoin: a ZerosWithPattern method to cuda_ndarray, and a slight refactoring of Join(Op) to inherit from this class in GpuJoin. And added GpuJoin itself. And a few unit tests for it.

numpy define the fct maximum and mininum as doing elemwise max and min.
We have the theano.tensor.{largest, smallest} that did that but with more op.
That make the graph harder to read. We will probably map largest and smallest
to the new fct in the futur after time show they are good.