Frame and FrameBatch improvements

[NicolasHug]

This PR:

• Adds a bunch of input checks when creating a Frame or a FrameBatch. We enforce Frame data to be 3D, and FrameBatch data to be >= 4D. We also ensure consistency of leading dimensions between data, pts_seconds and duration_seconds
• Adds indexing support for FrameBatch. This supports pytorch fancy indexing naturally and intuitively. Note that indexing a 4D FrameBatch returns a Frame.
• Adds iteration support for FrameBatch. This removes the "tuple unpacking" behavior that we had for FrameBatch, but luckily this is not something we have been using at all. The unpacking behavior of Frame is preserved.

These changes are mostly necessary in order for us to change the output of the samples from List[FrameBatch(4D)] to FrameBatch (5D), as done in #284

[ahmadsharif1]

Tensor has a .item() method for returning the underlying dtype.

Should we have something like that here? i.e. always return a FrameBatch but return a Frame if .item() is called?

[NicolasHug]

always return a FrameBatch but return a Frame if .item() is called?

I feel like this is what we're already doing, but perhaps I'm misunderstanding?

BTW, this quirk is only needed for mypy (sigh). Originally the code was simpler:

        cls = Frame if self.data.ndim == 4 else FrameBatch
        return cls(
            self.data[key],
            self.pts_seconds[key],
            self.duration_seconds[key],
        )

and everything was fine, and the Frame would get proper float value because of what we do in its post_init. But mypy was complaining so I had to go for this in 4661237 (#283)

[ahmadsharif1]

I feel like this is what we're already doing, but perhaps I'm misunderstanding?

Don't we return a Frame for the special case of dimensions=4?

What I am saying is we should return a FrameBatch even in that case (of size 1). So we are consistent with Tensor

I'll leave it to you

[NicolasHug]

I don't have a super strong preference on this - let me open an issue so we can discuss during one of the meetings