detect at run time the gpu pointeur size and int size. Use that to allow fusing more gpu elemwise together

Added error messages in cuda_ndarray, and a (commented) way of printing freed memory when debugging memory allocations (in the same file).

+++ theano/tensor/tests/test_basic.py (reindented)
@@ -1852,7 +1852,7 @@
             r = numpy.asarray([0.,1.,-1.], dtype=dtype)
             v = fn(l, r)
             self.failUnless(numpy.all(v == (l == r)), (v, (l==r)))
-
+
     def test_neq(self):
         for dtype in ['float64', 'float32', 'complex64', 'complex128']:
             x, y = vector(dtype=dtype), vector(dtype=dtype)

as they make a difference in the hash. They don't change the
generated c code for now.

This is a follow up from James code-review comment.