…torch

The new `TaskDataset` class provides conveniences like per-frame
annotations, bulk data downloading, and caching without forcing a dependency
on PyTorch (and somewhat awkwardly conforming to the PyTorch dataset
interface). It also provides a few extra niceties, like easy access to labels
and original frame numbers.

Note that it's called `TaskDataset` rather than `TaskVisionDataset`, as my plan is
to keep it domain-agnostic. The `MediaElement` class is extensible, and we can add,
for example, support for point clouds, by adding another `load_*` method.

There is currently no `ProjectDataset` equivalent, although one could (and
probably should) be added later. If we add one, we should probably also add
a `task_id` field to `Sample`.