(fix): handle none/dict rewards in _compute_zero_std_metrics

[atoniolo76]

In OPD or multi-turn cases, the sample.reward can come back as None or a dict of teacher log probs.

The _compute_zero_std_metrics function returns 1 if all samples in a group have the same reward, 0 otherwise. If a sample is aborted (common in multi-turn training) and sample.reward is None, Slime fails with TypeError: type NoneType doesn't define __round__ method at the round() call in interesting_rewards. For OPD, the sample.reward may be a custom dict containing a payload of teacher log probs for post_process to handle. This function intercepts the reward and fails with the same TypeError: type dict doesn't define __round__ method. Note that the original failure mode only happens when a group of rewards are all None or contain a dict. Now, when any rewards are None or dict, _is_zero_std is not run for the whole group.

To reproduce:

from argparse import Namespace
from slime.ray.rollout import _compute_zero_std_metrics
from slime.utils.types import Sample
 
args = Namespace(advantage_estimator="grpo", reward_key=None)
samples = [Sample(group_index=0, reward=None), Sample(group_index=0, reward=None)]
_compute_zero_std_metrics(args, samples) 

Will fail with # TypeError: type NoneType doesn't define __round__ method