Fixes #4363.
Fixes #4364.

When one creates an ndarray with np.longlong type, calling .dtype on the ndarray prints np.int64. As far as I can tell, the np.longlong type is equivalent to np.int64.

However, in C code, the NPY_LONGLONG macro has a different value from the NPY_INT64 macro.

This makes it so that we treat NPY_LONGLONG the same way as NPY_INT64. Alternatively, to fix these issues, I could reject NPY_LONGLONG and include better error messages but as far as I can tell np.longlong and np.int64 are the same.

Test Plan

Added a unit test case for creating tensors with np.longlong

Run the following (each script is from a separate issue) on python 2.7 and assert output is correct:

import torch
import numpy as np
any_number = long(2)
a = np.arange(0, any_number)
print(a)
b = torch.Tensor(a)
print(b)

import torch
import numpy as np
any_number = long(2)
a = np.arange(0, any_number)
print(a, a.dtype)
b = torch.from_numpy(a)
print(b)