Preserve NumPy sample values across non-native byte orders

[antonsoo]

NumPy arrays with a non-native byte order are passed to libsndfile as native samples, silently changing their values.

Convert write inputs to contiguous native storage in write(). Reject non-native read(out=...) buffers in _array_io, alongside the existing contiguity validation, before altering the output or moving the file position. Document the output-buffer restriction.

import io
import sys
import numpy as np
import soundfile as sf

order = ">" if sys.byteorder == "little" else "<"
x = np.array([1, 256, -2345], dtype=order + "i2")
file = io.BytesIO()
sf.write(file, x, 8000, format="WAV", subtype="PCM_16")
file.seek(0)
print(sf.read(file, dtype="int16")[0])

Before on a little-endian host: [256, 1, -10250]. After: [1, 256, -2345].

All 32 regressions fail on master and pass here: sixteen contiguous/strided non-native write cases and sixteen read-buffer rejection checks covering four sample dtypes, mono/stereo and fill settings. Recorded LJSpeech/AMI writes correct 123,616 sample values and match independent frame energies/spectra. All eight non-native output-buffer workflows reject before changing their buffers; all eight native controls remain byte-identical. Inputs remain unchanged.

Full native suite: 371 passed on CPython 3.12 and PyPy 3.10. Error lint and diff checks pass. Pyright retains the same two existing dirname(_libname) findings as master. Tested on little-endian Linux; no big-endian host was tested.

[bastibe]

Thank you for your bug report. Were you actually using soundfile on a big endian system?

I don't think this is the right place to fix things, though. We should probably treat byte order like contiguous instead.

[bastibe]

here we convert data to contiguous, we could also convert the byte order.

[antonsoo]

Moved native byte-order conversion to write(), alongside its contiguity conversion. Contiguous and strided inputs are covered; all sixteen write regressions pass on CPython and PyPy.

[bastibe]

if out is not contiguous, we just error out. We should do the same with byte order.

[antonsoo]

Non-native output arrays now raise ValueError before any samples are read. Tests verify the output bytes and file position stay unchanged, including fill settings.

[antonsoo]

The reproduction used non-native-endian NumPy arrays on little-endian Linux; no big-endian host was tested.

Applied your suggested layout: write() now converts byte order alongside contiguity, and read(out=...) rejects non-native buffers before changing the buffer or advancing the file position. Removed the conversion/copyback recursion from _array_io and documented the output-buffer restriction.