[RLlib][early-kickoff] RNNs and RLModules

[ArturNiederfahrenhorst]

Note: This PR started as a fork of #30997 to experiment with how we incorporate RNNs into RLModules.

Why are these changes needed?

This PR makes stateful RLModules compatible with our legacy sampling stack.
Most notably, this includes the following changes:

• SampleBatches in the RNN case are pushed out of the EnvRunner with tensors in [B, T, ...] format.
• We add Tokenizers. Tokenizers are simply Encoders that live inside an LSTM encoder or GRU encoder.
• RLModules update Policy's view requirements (similar to what Models did before).
• RLModules get a is_recurrent method.
• A method auto_fold_unfold_time that our heads use under the hood to automatically reshape inputs that have a time dimension before pushing them through the model and reshaping them back to add back the time dimension afterwards.

Additionally, this PR deals a lot with our new place for states "state_out" vs the old "state_out_..." and makes places distinguish between the RLModules and non-RLModule case.

[kouroshHakha]

So there are two quality issues with this PR which I am intentionally postponing to another follow-up / clean up PR after the branch cut for 2.6:

1. Having the auto_fold_unfold_time decorator depend on the spec checking module is not the cleanest solution. Also there is a lot of reduntant code between tf / torch which would make maintenance a huge pain. @ArturNiederfahrenhorst We should come back and fix this issue
2. The overriding of v and r values in compute_advantages does not sound right. In my opinion the caller should write the correct v and r inside the SampleBatch already. We should investigate that further.