Downgrade torch to 2.5.1 or lower to avoid CPU memory leaks

[StoneT2000]

We recently find that using torch 2.6.0 can lead to CPU memory use growing endlessly. Particularly it occurs when we try to transfer GPU memory from the physics engine to a torch tensor on device. It is still unclear what the true cause of it is but it looks like something on PyTorch's end. Downgrading to 2.5.1 or lower fixes this.