[BUG] Segfault or wrong value on array with non-vanilla strides.

[lusewell]

Describe the bug
Segfault or wrong answer when running various functions on a numpy array with unusual strides. Looks like reading outside the array.

To Reproduce

import bottleneck
import numpy as np
data = np.zeros((1,500, 2)).transpose(1,2,0)
bottleneck.nanmax(data) # This inconsistently either segfaults or gives a non-zero very large number.

bottleneck = 1.3.2
numpy = 1.16.6
python = 3.7.7

Operating system is RHEL7.

I've also replicated on ubuntu with python 3.9 and numpy 1.20, (still bottleneck 1.3.2) so doesn't seem to be too sensitive to versions of other packages.

Expected behavior
The nanmax of this data should always be zero.

Additional context
This in turn breaks xarray.DataArray.max

[lusewell]

I believe the bug comes from here assuming that the strides are sorted.