Turns out I was testing from the wrong build — oops. That explains why the format mysteriously stopped being NV12. Sorry for wasting everyone's time.
When I actually run the correct version, we get NV12-formatted data in `evr_copy_sample_buffer`, and hence just the luma plane, resulting in a wonderful monochromatic green image. Copying the extra plane data then fixes it.
I've pushed a hacky implementation of this (we treat the extra plane as just being more scanlines for the first plane. This works, as they're guaranteed to be the same width / pitch, but I can use a second `MFCopyImage` call instead if that's preferred.
I suspect the _right_ way of doing this involves more explicitly supporting 'multi-plane formats' and looping over them, but a hacky 'if (format == NV12) MFCopyImage(2nd plane)' could work if we don't want to overcomplicate this while there's only one plane.
(I also made the FIXME() an error, which was my original intention with the fallback. I'm happy to just assume we'll never get an unsupported format, though.)