Minimum torch version

[bsprenger]

torchRL currently pins the lower bound for the torch version to a very-old 2.1.0. As discussed here, tensordict also does not pin a lower bound for torch, and is considering pinning 2.13.0 as a lower bound.

Accordingly, we are considering updating the torch version lower bound to 2.13.0 for torchRL, for the reasons described in the linked thread. (Note that if tensordict pins the min version to 2.13.0, then even if this lower bound is set to 2.1.0 here, it would effectively be pinned at 2.13.0 since tensordict is a direct dependency of torchRL).

Since this could affect lots of users, we wanted to get your feedback:

• Is there a good reason to support such an old version of torch that we are missing?
• Is there any widely used downstream libraries which require versions before 2.13.0 that we should consider before pinning the lower bound to 2.13? Would those libraries' use cases be satisfied with simply using an older version of torchRL?

Curious to hear everyone's thoughts!

@theap06 , whenever you get time later this week to run the CI test matrix with different torch versions, that would be great info too.

cc @peterdsharpe

[theap06]

@bsprenger I had an agent create a bash script and run the full test suite. I'm also ok with using 2.13 as the lower bound.