Workplan for v0.4.0 (WIP)

[vmoens]

In the next release, we'd like to incorporate the following new features:

## Objectives

• Have all objectives work with native nn.Modules: We wish to be able to get all modules to work as in DQNLoss(nn.Linear(3, 4)) where the linear module is your value network. It's unclear how this will work when a distribution is needed (eg PPO). We will also have to overload the forward to patch it with the signature provided by @tensordict.dispatch and ensure compatibility with torch.compile. We also want to make this a blessed feature, presented clearly and concisely in the doc.
• Fix dreamer @BY571
• Fix REDQ
• Add reduction methods to all losses (eg, mean, max...) @albertbou92
• Overload forwards in losses [NOMERG] Add @overload to forward in losses #1893
• Use tensorclass outputs in losses [NOMERG] Prototyping tensorclass for losses #1892

Data

• Add more components for efficiently work with the replay buffers, such as [Feature Request] PrioritizedSliceSampler #1876 or preprocessing for datasets (cc @nicklashansen).

General

• Better compatibility with MPS