fix(rollout): reject aborted samples before reward tensor

[YeonwooSung]

Fixes #200.

Aborted samples keep reward=None. On the default GRPO path that list was passed to torch.tensor, which crashes. Custom generate functions that abort (e.g. Search-R1 tool failure) hit this.

This PR fail-closes instead of scoring those samples. We do not default reward to 0.0. A previous attempt that did (#1252) was closed unmerged; aborted / unrewarded samples must not reach training.

Change

In RolloutManager._post_process_rewards, on the default path only (when custom_reward_post_process_func is None), reject any sample with status == Sample.Status.ABORTED or reward is None before get_reward_value / torch.tensor.

The error tells callers to filter in the generate function or recycle via partial rollout. Custom post-process keeps its own contract and is not subject to this check.

get_reward_value is self.reward or self.reward[args.reward_key]. Checking reward is None first avoids a TypeError when reward_key is set.

Tests

tests/test_post_process_rewards.py (NUM_GPUS = 0) binds _post_process_rewards onto a dummy object (no Ray):

1. Two completed numeric rewards succeed and return tensor-compatible lists
2. One ABORTED sample raises ValueError
3. COMPLETED with reward is None raises ValueError
4. A custom post-process function is used as-is (aborted sample does not hit the new check)