Today the mjx_env.init and mjx_env.make_data functions don't take in the device which mjx is supposed to use and this makes mjx use the default device which is cuda:0. This is problematic in pytorch ...