[BugFix] Clone memmap tensors on regular tensors and other replay buffer improvements

[vmoens]

Description

This PR introduces several changes aimed at making replay buffer sampling more efficient:

• MemmapTensor.clone() now returns a regular tensor. This is a more intuitive behaviour than the previous one (where another MemmapTensor was returned). Also, to get another MemmapTensor, one can simply call MemmapTensor(memmap.clone()).
• tensordict indexing now never returns a SubTensorDict. To get a SubTensorDict, one must built it explicitely by calling the appropriate dedicated method. This is to make sure that no overhead is introduced by working with this class.
• Replay buffer storage is now cast to the appropriate device by the helper function: this makes data sampling faster and avoids the creation of tensors on cpu that are then cast on gpu (tensors are now created on gpu immediately).

[vmoens]

REDQLoss is memory intensive when used with large models.
Some work is needed to make it less memory intensive.