Add reference to policy with state dict

[felixy12]

Description

This PR adds the _policy_w_state_dict attribute for the SyncDataCollector. The point here is to use this reference to do all saving/loading of the policy state dict. This fixes the bug of not being able to load/save policies that have been compiled/CudaGraph'ed.

Motivation and Context

Policy models are often saved and loaded via the SyncDataCollector class, which also handles compiling and CudaGraphing the policy. However, this wraps the policy in a few layers that obscures the saving and loading of the policy weights if we attempt the access the policy directly from the SyncDataCollector.policy attribute.

The clean solution here is to keep a reference to the unwrapped, original policy object that contains the state_dict and perform all model saving/loading via this reference. This way, we don't care about the method in which the object gets wrapped. This all works because every wrapped layer still points to the same underlying set of model weights.

• I have raised an issue to propose this change (required for new features and bug fixes)

Types of changes

What types of changes does your code introduce? Remove all that do not apply:

• Bug fix (non-breaking change which fixes an issue)

Checklist

Go over all the following points, and put an x in all the boxes that apply.
If you are unsure about any of these, don't hesitate to ask. We are here to help!

• I have read the CONTRIBUTION guide (required)
• My change requires a change to the documentation.
• I have updated the tests accordingly (required for a bug fix or a new feature).
• I have updated the documentation accordingly.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/3043

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.