fix: correct generated group type annotations

[guapisolo]

Summary

Correct generated rollout group annotations to include expanded sample lists.

Symptom & Reproduction

• Symptom: generate_and_rm_group declares a flat list[Sample] even though one trajectory can expand into list[Sample], so type consumers receive a false flat-only contract.
• Reproduction: Inspect typing.get_type_hints(generate_and_rm_group)["return"] while tracing the asyncio.gather inputs from generate_and_rm.

Root Cause

1. generate_and_rm returns Sample | list[Sample] per trajectory.
2. asyncio.gather preserves each trajectory result inside the group list.
3. generate_and_rm_group and its direct sync consumer declared only list[Sample].

Fix

Declare the producer result and the direct sync consumer as list[Sample | list[Sample]], matching the existing runtime shape without changing generation, ordering, reward, or filtering behavior.

Verification

• New / updated test: No permanent test was needed.
• Type contract: A focused get_type_hints() assertion verifies the corrected contract.
• Existing test suite: 30 rollout tests verify preserved behavior.

Review Focus

• Scrutinize the annotations in generate_and_rm_group and generate_rollout_async.
• Confirm the diff contains no runtime statement changes.