PR #458 causes tianshou 0.4.4 fails to load checkpoints that was previously trained on tianshou 0.4.3.

I think a simple conversion is needed. Post it here in case anyone needs it.

data = torch.load('previous.pth')
for k in list(data.keys()):
    data['_actor_critic.' + k] = data[k]
torch.save(data, 'new.pth')

Originally posted by @ultmaster in #458 (comment)