Large arrays casting of unicode/bytes cast to quaddtype leads segfault

[SwayamInSync]

import numpy as np
from numpy_quaddtype import QuadPrecDType

arr = np.arange(501, dtype=QuadPrecDType())
arr.astype('U50')

# output
# Segmentation fault (core dumped)

Interesting part is this happens as size > 500 i.e. segfaults for 501 and ahead
This seems more of a GIL issue as per docs functions like PyUnicode_FromString should be called with GIL held and this issue get resolved when we additionally register the loops with NPY_METH_REQUIRES_PYAPI flag.

Should we take this as a fix? I saw StringDType also uses this flag in casting spec

cc: @ngoldbaum @seberg

[seberg]

Ah, sorry, I think it semi-registered, but I forgot to point out/ensure it's the case. If you want to use that API (not quite sure how avoidable it is), then yes, you need to indicate NPY_METH_REQUIRES_PYAPI on in the method flags in the registration and all is fine.

the 500 threshold is just because releasing the GIL (or probably rather re-aquiring it) at least used to be quite slow in Python, so we avoid it. I think it may be faster on newer versions and it might be practical to just always release these days.

[SwayamInSync]

Avoiding API might need good enough changes from the dragon4 file (I can keep this in future todo)
For now we can keep the flag then, I will push a PR resolving both unicode and bytes with the the flag after merging of #228

[ngoldbaum]

I don't think it's problematic to require the GIL here. Printing and formatting numbers isn't normally performance-critical.