Replaces PyObject_Malloc/Free with their PyMem counterparts; PyPy does
not support the PyObject variants. Arguably, based on
https://docs.python.org/2.3/whatsnew/section-pymalloc.html the PyMem
variants are a better fit for this use case anyway (PyObject_ being
intended for "small" allocations).

Disables the use of CP1250 under PyPy because it lacks the
PyUnicode_Encode function.

There are no new tests failures. (testConnectWithWrongDB fails with a
1044 error for me and not 1049 with both the original code under
CPython and this code under PyPy).

If we are reading many packets of different sizes from a fast database
server, it's possible that the skip() method of the packet reader will
never find the start and end pointers aligned and so it will never reset
them. This can lead to throwing 'Socket receive buffer full' errors when
we reach the end, even if there's plenty of unused space at the start of
the buffer from previous packets.

This checks for this situation and moves the live data and current
pointers to the start of the buffer, thus making room (if possible).