Fix for reading certain Ogg files on Linux via file-like objects.

[psobot]

This PR fixes a subtle bug where certain Ogg Vorbis files can fail to decode when passed to Pedalboard via Python file-like objects, particularly on Linux.

The bug stems from an interaction between three layers of code:

1. When Pedalboard reads audio from a Python file-like object (i.e.: a network stream, custom wrapper, etc), it calls Python methods like read(), seek(), and tell().
2. These Python objects can internally call other native C/C++ code that sets the thread-local errno variable as part of normal operation, but this error is not enough to cause the Python wrapper(s) to throw exceptions.
3. The libvorbis decoder (specifically in vorbisfile.c) checks errno after I/O operations to detect errors:

   errno=0;  // Clear errno before the read
   // ... perform read operation ...
   if(bytes==0 && errno)return(-1);  // Treat errno!=0 as an error

This results in valid Ogg Vorbis files failing to parse, as - if that last libvorbis read returns 0 bytes, but some other code set errno - then the entire read call fails. This can be spurred on by appending a couple null bytes to the end of the file, which I do in test here.

[psobot]

This was the only reliable way I could find to repro the bug in test, apart from including a proprietary audio file as part of the tests. Trailing null bytes should not prevent the file from being parsed.

[hyperc54]

If it's supposed to trigger the bug but now this PR fixed it, should we actually test this with a simple

    with pedalboard.io.AudioFile(buf) as af:  # <- the buffer with garbage in it.. which should now succeed parsing  without any change
        assert af.samplerate == 44100  

and we can keep the other assert with a different but perfectly fine buffer

    with pedalboard.io.AudioFile(ErrnoTriggeringFileLike(ok_buffer_made_not_ok_with_wrapper_class)) as af:
        assert af.samplerate == 44100

[psobot]

Not sure I fully follow - the code here does fail on Linux if I comment out the errno = 0 fix in the code under test.

[hyperc54]

the code here does fail on Linux if I comment out the errno = 0 fix in the code under test.

This makes sense 👍

What I was challenging is this part:

    # Add some random garbage past the end of the file to trigger the bug:
    buf.write(b"\x00\x00")

I would argue you don't need to add random garbage because the test buffer is wrapped in ErrnoTriggeringFileLike which would trigger the bug in any case with any buffer, clean or not. (If I understand the class properly)

I was then suggesting to actually test without the ErrnoTriggeringFileLike and an unhealthy buffer :)

not important though!

[hyperc54]

Great dig and explanation, thank you

[hyperc54]

If it's supposed to trigger the bug but now this PR fixed it, should we actually test this with a simple

    with pedalboard.io.AudioFile(buf) as af:  # <- the buffer with garbage in it.. which should now succeed parsing  without any change
        assert af.samplerate == 44100  

and we can keep the other assert with a different but perfectly fine buffer

    with pedalboard.io.AudioFile(ErrnoTriggeringFileLike(ok_buffer_made_not_ok_with_wrapper_class)) as af:
        assert af.samplerate == 44100