In-place `Quantity` operations with scalars and arrays can cause segmentation faults with `numpy`>=1.21

Description

Trying to add in-place an array Quantity to a scalar Quantity should produce a ValueError (but see also #10072). However, with numpy>=1.21 it produces a segmentation fault, at least on Linux.

Suppose the following code snippet:

>>> from astropy import units as u
>>> scalar = 1*u.m
>>> array = (2, 3)*u.m
>>> scalar += array

Behavior with numpy<1.21

With astropy 5.0.4 (LTS) and numpy<1.21:

ValueError: non-broadcastable output operand with shape () doesn't match the broadcast shape (2,)

Behavior with numpy>=1.21

With astropy 5.0.4 (LTS) and numpy>=1.21:

Segmentation fault (core dumped)

I have verified that the segmentation fault also happens with astropy main (2288ecd4e9c4d3722d72b7f4a6555a34f4f04fc7) and numpy 1.22.3.

[eerovaher]

Given that the segmentation fault indeed only occurs when the operation that is being attempted is invalid to begin with, I now agree that there is no need to address this any further.

[mhvk]

Yes, indeed, let’s close this. Thanks for reporting, @eerovaher - now numpy is a bit better again!