[Bug] Fully-async multi-agent rollout crashes on nested groups

Bug Description

The fully-async multi-agent demo crashes when a custom generation function returns multiple samples for one input sample.

Affected configuration in examples/multi_agent/run-qwen3-8B-multi-agent.sh:

--rollout-function-path slime.rollout.fully_async_rollout.generate_rollout_fully_async
--custom-generate-function-path examples.multi_agent.rollout_with_multi_agents.generate_with_multi_agents

The custom generation result becomes a nested group (list[list[Sample]]), but fully_async_rollout.py expects a flat group and treats Python's list.index method as Sample.index.

Steps to Reproduce

cd slime
bash examples/multi_agent/run-qwen3-8B-multi-agent.sh

The relevant settings are:

--rollout-batch-size 16
--n-samples-per-prompt 4
--sglang-server-concurrency 16

A minimal reproduction is:

from slime.utils.types import Sample

nested_group = [[Sample(index=7)]]

for item in nested_group:
    index = getattr(item, "index", None)
    if index is not None:
        int(index)

Expected Behavior

The rollout should handle nested groups and continue training. It should find the leaf Sample.index and sort groups by that value.

Actual Behavior

The rollout crashes while sorting completed groups with:

TypeError: int() argument must be a string, a bytes-like object or a real number, not 'builtin_function_or_method'

For a nested group, getattr(item, "index", None) returns Python's built-in list.index method instead of a sample index.

Environment

- Python: 3.12.13
- PyTorch: 2.9.1+cu128
- CUDA: 12.8
- GPU: NVIDIA H20, 8 GPUs
- OS: Linux
- SGLang: 0.5.10.post1

Logs

Additional Context

No response

Pre-submission Checklist

• I have read the CONTRIBUTING.md and understand the collaboration scope.
• I have read the documentation and my issue is not addressed there.
• I have searched for existing issues and this is not a duplicate.
• I have provided a minimal, reproducible example.

Start with examples/multi_agent/run-qwen3-8B-multi-agent.sh and the fully_async_rollout.py sorting path named in the report, then compare its expected group shape with examples/multi_agent/rollout_with_multi_agents.generate_with_multi_agents. Reproduce with the provided nested Sample example or the demo command. Done means nested groups locate the leaf Sample.index, sort successfully, and the multi-agent rollout continues training.