numpy.load cannot read from tar archive in python3 ('_FileInFile' object has no attribute 'fileno')

I would like to numpy.load an ndarray from a file within a tar archive, without extracting to disk, in Python3.

import os
import numpy
import tarfile

# create a test tar archive
a = numpy.random.rand(10, 4)
numpy.save('a.npy', a)
with open('foo.txt', 'w') as f:
    f.write('hey\nho\n')
t = tarfile.open('abc.tar', 'w')
t.add('a.npy')
t.add('foo.txt')
t.close()
del t, a
os.remove('a.npy')
os.remove('foo.txt')

# -----

t = tarfile.open('abc.tar', 'r')
a = numpy.load(t.extractfile('a.npy'))

The above code snippet works fine in Python 2.7.12 with numpy 1.11.1, but fails in Python 3.5.2 with numpy 1.11.1 (Debian Linux):

Traceback (most recent call last):
  File "loadtest.py", line 21, in <module>
    a = numpy.load(t.extractfile('a.npy'))
  File "[...]/lib/python3.5/site-packages/numpy/lib/npyio.py", line 406, in load
    pickle_kwargs=pickle_kwargs)
  File "[...]/lib/python3.5/site-packages/numpy/lib/format.py", line 648, in read_array
    array = numpy.fromfile(fp, dtype=dtype, count=count)
AttributeError: '_FileInFile' object has no attribute 'fileno'

Am I doing something wrong? The numpy docs say that I can pass a file-like object to numpy.load, which must support the seek() and read() methods, which the io.BufferedReader I receive from TarFile.extractfile() does.

Start with the reproducer in the issue and inspect numpy/lib/npyio.py and numpy/lib/format.py at the failing numpy.load and numpy.fromfile calls. Run the Python 3 tarfile example first, then verify that loading the array from TarFile.extractfile() succeeds without extracting it to disk.