test: assert that isinf returns -1 for negative infinity

Fixes for buffer protocol docs

[ci skip]

Makes sure (and documents, in a sense) that isinf's declaration as returning
a bint preserves its C semantics.

This will be useful for, e.g., implementing a fake cythonize that
can leverage the deductions made by the real cythonize, but could
have other uses as well.

Their temporaries can't be references.

Docs: how to implement PEP 3118, the buffer protocol

work around memoryview.tobytes() off-by-one bug with suboffsets

In CPython 2.7, the memoryview.tobytes() method (implemented in
function memoryview_tobytes() in cpython/Objects/memoryobject.c) calls
PyBuffer_ToContiguous() to copy the bytes in the buffer object to the
newly allocated string memory.  PyBuffer_ToContiguous() in turn calls
PyBuffer_IsContiguous(), which always returns false when the
suboffsets member of the Py_buffer struct is non-NULL (even if all
entries in that array are negative!).  When PyBuffer_IsContiguous()
returns false, PyBuffer_ToContiguous() runs an alternative memory copy
scheme which apparently has an off-by-one bug that the normal memory
copy scheme doesn't have.

This change sets the Py_buffer suboffsets member to NULL if all
entries are negative, avoiding the off-by-one bug.

To reproduce the bug:

    cpdef foo():
        cdef unsigned char[:] v = bytearray("testing")
        # the following prints 'estingt' without this workaround
        print repr(memoryview(v).tobytes())
Signed-off-by: Richard Hansen <rhansen@bbn.com>

In CPython 2.7, the memoryview.tobytes() method (implemented in
function memoryview_tobytes() in cpython/Objects/memoryobject.c) calls
PyBuffer_ToContiguous() to copy the bytes in the buffer object to the
newly allocated string memory.  PyBuffer_ToContiguous() in turn calls
PyBuffer_IsContiguous(), which always returns false when the
suboffsets member of the Py_buffer struct is non-NULL (even if all
entries in that array are negative!).  When PyBuffer_IsContiguous()
returns false, PyBuffer_ToContiguous() runs an alternative memory copy
scheme which apparently has an off-by-one bug that the normal memory
copy scheme doesn't have.

This change sets the Py_buffer suboffsets member to NULL if all
entries are negative, avoiding the off-by-one bug.

To reproduce the bug:

    cpdef foo():
        cdef unsigned char[:] v = bytearray("testing")
        # the following prints 'estingt' without this workaround
        print repr(memoryview(v).tobytes())
Signed-off-by: Richard Hansen <rhansen@bbn.com>

--HG--
extra : transplant_source : h%F68%C5f%D2%A9%86%A7o%B4w%27%E5%EDt%3D%D5%3B%3C

Add some declarations for PyTypeObject

--HG--
extra : transplant_source : %05%8F%EB%B6%98%1E%DF%ED%21%84%D0%0B%05m%AB%D0%97%27%09%1D

fix erroneous prototypes for npy_logaddexp*