This moves the lammps.get_neighlist() method to lammps.numpy.get_neighlist().
lammps.get_neighlist() now returns a NeighList object, while the NumPy variants
returns a NumPyNeighList object. The main difference between the two is that while
the ctypes variant returns neighlist elements as
atom idx (int), numneighs (int), neighbors (POINTER(c_int))
the NumPy variant returns
atom idx (int), neighbors (numpy.array)
LammpsNumpyWrapper was a class that was defined inside of the
lammps.numpy property when it was first accessed. This made it hard
to document the methods of this class.
This commit extracts this utility class into the lammps module and
renames it to 'numpy_wrapper'.
Keep LAMMPS_TAGINT and LAMMPS_BIGINT internal to LAMMPS.
An external caller now only needs to distinguish between 32 and 64bit
integers, doubles and C strings.
Both extract methods now can auto-detect the datatype of both global
and per-atom properties. Callers can still enforce different types
if needed by specifying the now optional dtype argument.
The numpy wrapper now has a new extract_atom function method, which
replace the extract_atom_darray and extract_atom_iarray method and
autodetects both type and size. All parameters can still be forced
to use different values if needed.
Fixes segfaults caused by API change. The API change in
lammps_open and lammps_open_no_mpi makes them return the LAMMPS pointer
via their return value. However due to how ctypes operates, even
if restype is specified to be c_void_p, the function returns an integer.
Without the proper type of the pointer, calling functions without arglists would default
to using 32bit integers to pass an argument, which cuts away parts of the 64bit pointer.
Subsequently, resolving the truncated pointer in the library causes segfaults.
This commit fixes the root cause. But it also highlights the need of specifying
the arglists of all library functions.
