Don't borrow __getitem__ from numpy

[JBlaschke]

Currently we use numpy.array.__getitem__ on the host to do the indexing operation -- this is a bit inefficient was it requires DtoH and HtoD Memcpys.