[RLlib]: Fix torch None conversion

[Rohan138]

PyTorch does not support dtype=object anymore, which is what np.asarray(None) returns. This breaks save_checkpoint and load_checkpoint for algorithms with pytorch >= 1.12.0.

Repro script:

from ray.rllib.algorithms.dqn import DQNConfig
import os
import shutil

checkpoint_dir = "/tmp/cartpole/torch/"
shutil.rmtree(checkpoint_dir, ignore_errors=True)
os.makedirs(checkpoint_dir, exist_ok=True)
config = (
    DQNConfig()
    .environment(env="CartPole-v0")
    .framework("torch")
    .rollouts(num_rollout_workers=4)
)
algo = config.build()
random = algo.save_checkpoint(checkpoint_dir)
algo.load_checkpoint(random)

Related issue number

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[avnishn]

This is fine with me pending tests passing. Thanks for submitting this!