Extend `SimpleVideoDecoder` index-based APIs

[NicolasHug]

For consistency with tensor indexing semantics and with builtin Python indexing of sequences, we should extend the index-based APIs of SimpleVideoDecoder, namely __getitem__(), get_frame_at() and get_frames_at() to support:

• negative indexing. -n just means len(decoder) - n.
• upper bound in slices to be greated than len(decoder).

In short we should allow for that kind of stuff:

>>> torch.arange(10)[-3:1000]
tensor([7, 8, 9])

In terms of implementation, we can implement this within the SimpleVideoDecoder class in Python. No need to extend the core or C++ APIs.