Feature Request: Read from Bytes

[RyougiKukoc]

In terms of huge video dataset NN training, it is common to use DataBase like LMDB or webdataset to packing scattered images or videos togather, automatically mapping to memory in DDP and multi-async-workers training pipelines.

Well, the problem is, this paradigm can't avoid transfering a file object into its equivalent bytes representation, which is not currently supported by any source filter in VapourSynth as far as I know. The community has made such huge progress in parsing various kind of AV formats that any backend that commonly used in DL cannot satisfy me. It is common to see the inaccurate metadata parsing or inaccurate frame seeking here and there (pyav, opencv, torchvision). So I turn to pull up this feature request.