ndarray: import python standard library complex arrays

[hpkfft]

Python 3.15 adds support for complex arrays to the standard library's array module. It uses a single capital letter as the format code. This PR adds support to nanobind's ndarray to import these arrays using the buffer protocol.

I tested this locally. The array support will be in Python 3.15 alpha 8. When testing nanobind on older releases, the new test is skipped.

PEP 3118 proposed additions to the struct string syntax which were not added. One of these was that Z followed by another character describes a complex number. For example, 'Zf' describes a complex single-precision number. For example, this PR allows nanobind to import complex single precision arrays via the buffer protocol that use either 'Zf' or 'F' as the format code. (Also, 'ZF' is accepted, but that's unintentional.)

[hpkfft]

Note that we export arrays using the buffer protocol using the format proposed by PEP 3118 because that's what NumPy requires for import. NumPy cannot import buffers using 'F' or 'D'. (NumPy has half-precision reals, but not complex, so complex32 is currently irrelevant to them.)
Consequently, nanobind cannot export complex ndarrays as Python standard library arrays.
(Neither can NumPy. NumPy cannot import them either, but we can at least do that.)

[hpkfft]

@ngoldbaum, if you have time, I'd appreciate having your thoughts on a related Python question. Above, format is set to a string literal. I would guess doing this is common. The compiler/linker puts these constant strings in read-only memory. In the Python header file pybuffer.h, this is char *format;. Would it be better for Python to specify pointer-to-const, i.e., const char *format;?

[wjakob]

Thanks!

[hpkfft]

For those interested, the Python Steering Council may offer their opinion: python/steering-council#345