In NumPy 1.16.0, np.load() fails to load some files created by np.save(), saying "ValueError: invalid shape in fixed-type tuple."

The problem seems to be that #12358 changed np.lib.format._read_array_header(). It used to pass the descr to np.dtype(), but now it uses a new function descr_to_dtype() which fails for some valid inputs.

We have significant amounts of data stored on disk with formats which NumPy 1.16.0 cannot load, so this is a showstopper for upgrading beyond 1.15.

Reproducing code example:

This fails in NumPy 1.16.0, but works in all prior versions:

import numpy as np
a = np.zeros(1, np.dtype([('c', ('<f8', (5,)), (2,))]))
np.save('a.npy', a)
np.load('a.npy')

Error message:

Traceback (most recent call last):
  File "numpy/lib/npyio.py", line 440, in load
    pickle_kwargs=pickle_kwargs)
  File "numpy/lib/format.py", line 678, in read_array
    shape, fortran_order, dtype = _read_array_header(fp, version)
  File "numpy/lib/format.py", line 563, in _read_array_header
    dtype = descr_to_dtype(d['descr'])
  File "numpy/lib/format.py", line 282, in descr_to_dtype
    dt = numpy.dtype((descr_to_dtype(descr_str), shape))
  File "numpy/lib/format.py", line 282, in descr_to_dtype
    dt = numpy.dtype((descr_to_dtype(descr_str), shape))
ValueError: invalid shape in fixed-type tuple.

Numpy/Python version information:

NumPy 1.16.0
Python 3.6.6