All the checks will still be applied in the inner function at least once
on every input value. However, for subsequent calls of the outer
node on the same input values (for output memory checking), the inner
function call will be much faster.

That way, some instances of DebugMode can do the checks,
while others do not.

Since the contract regarding output memory has changed,
it is not guaranteed that the op itself did allocate that memory.

Also, create the memory buffers as needed.

Fixed crash in a specific concatenate situation

Mixed

Fixed bug with broadcastable flags of MaxAndArgmax

Also add test for the mul grad.

Fix a number of tests for Python 2.4

int(numpy.bool_(False)) returned False, not 0.

Also add (disabled by default) debug printing statements.

Mixed

When linking with libpython 2.4 (at least), we cannot Py_INCREF /
Py_XDECREF PyObject pointers that have been declared constant.

Add '-m64' to the compilation line under win64

Better error msg and test reshape in the not contiguous case.