The bug make the CudaNdarray.__getitem__ return something with the wrong number of dimensions.

The bug happen with the param of __getitem__ is a tuple with element that are a numpy.int under Ubuntu(system where PyInt_Check(numpy.int) return false.

Their is many different rounding algo. So we don't create a round fct, but one for a 2 differents version(c round fct() and numpy.round)

We do this as in some case the test was failing during theano fct call.(i.e. under ubuntu)
As we use theano.tensor.scalar and not theano.scalar.scalar normally this is not a problem.

added comment to scalar c_extract function, explaining one way to go about fixing the broken scalar module

straight through to the compiler call. It affects the Op signature if it is not
empty.

- When extracting num and denum constants, wrap them in T.Constants,
so they keep their dtype, and cannot be downcasted
- Upcast all elements in num and denum to get the type of the neutral element,
if it is not specified
- Cosmetic changes

Fix a bug found with new test. We did not raise correctly error when we where out of bound.

Added more tests case that cover those case.

sparse - disabled a test that makes no sense since disabling the structured_dot_csc optimization a few days ago

The basic idea is to apply autocasting to constants but not to shared scalars.

Joint work with Fred.