Add support for numpy buffer format `Zf` and `Zd`

[vstinner]

• Python 3.14 added D and F formats to the ctypes and struct module
• Python 3.15 adds D and F formats to the array module and to memoryview

The D and F formats were chosen for compatibility with numpy. Problem: in fact, numpy uses buffer formats Zd and Zf. Example:

>>> import numpy as np
>>> memoryview(np.array([1, 2, 3], dtype='F')).format
'Zf'
>>> memoryview(np.array([1, 2, 3], dtype='D')).format
'Zd'

So I propose adding support for Zd and Zf formats in array, ctypes, memoryview and struct. It requires changing the code expecting a single character to now accept a string (Zd and Zf have 2 characters!).

Linked PRs

• gh-148675: Add Zd/Zf formats to array, ctypes, memoryview, struct #148676
• gh-148675: Revert PR 146237 and PR 146241 #148674

[vstinner]

This issue was escalated to the Steering Council: Decide buffer protocol format string for complex numbers.

[skirpichev]

The D and F formats were chosen for #132827. Problem: in fact, numpy uses buffer formats Zd and Zf.

In fact, numpy uses both. So, 'F'/'D' codes does make sense for the ctypes module, at least. If SC approves the compatibility break - we can change buffer formats to Zd/Zf in the ctypes module.

So I propose adding support for Zd and Zf formats in array, ctypes, memoryview and struct. It requires changing the code expecting a single character to now accept a string (Zd and Zf have 2 characters!).

I think it's better to revert changes for array and memoryview (#148674) and leave problem for a new champion, who will make a PEP with a proper extension of the struct format string for new types.