[Feature Request] cloning inconsistencies

Motivation

clone() should either return an object of the same type (e.g. SavedTensorDict.clone() returns a SavedTensorDict instance) or a generic type (e.g. MemmapTensor.clone() returns a torch.Tensor instance.)

Pros of same-type

• It makes sense that cloning an object returns a copy of this object (not another type)
• In the case of multi-task, we use the feature that a LazyStackedTensorDict contains tensordicts whose keys can be hidden. However, if cloning returns a regular TensorDict, those keys will be lost.

### Pros of generic type

• For users that do not have a very deep understanding of the repo's primitives, it can make more sense for clone to return the generic type. Also, if cloning a stack of tensordicts returns another stack of other tensordicts, it is hard to see what the advantage and what the usage could be.