Add ops test for file-like seeks

[scotts]

Additional test for file-like behavior; see test comments for more.

[NicolasHug]

We could store the number of samples of each frames in the .json files that we check in. Not sure that would be enough if the number of reads still depends on the internal FFmpeg buffers, as you mentioned.

[scotts]

I'm not even sure that would be enough to calculate the number of bytes decoded for each frame; I think we'd be going down the path of tracking the encoded data itself, which I don't think we should do. And that still wouldn't get us the determinism we need, as there's the internal buffers.

[NicolasHug]

Just noting this so we don't forget, we should test for proper errors if the user provides a malformed object that e.g. doesn't support read() or seek() correctly. E.g. if read() doesn't exist, right now we only error:

    def create_from_file_like(
        file_like: Union[io.RawIOBase, io.BytesIO], seek_mode: Optional[str] = None
    ) -> torch.Tensor:
        assert _pybind_ops is not None
>       return _convert_to_tensor(_pybind_ops.create_from_file_like(file_like, seek_mode))
E       NotImplementedError

which isn't super informative. I guess the check can be done on the public Pythons side rather than on the ops?

Also, we should try to hit some of the error checks, e.g. catching in the tests that we properly raise https://github.com/pytorch/torchcodec/blob/f416dcf8cb709d679cd739d2a56a761d8658468f/src/torchcodec/decoders/_core/AVIOFileLikeContext.cpp#L53-L59 when doing something bad like

return self._file.read(size + 10)

[scotts]

Excellent point. I'll put up another PR with such a test.