fix(fully-async): respect partial_rollout=False when requeuing ABORTED groups

[Kagura-0001]

Problem

When using fully_async_rollout with partial_rollout=False, ABORTED sample groups are unconditionally requeued back to data_buffer, bypassing the partial_rollout flag entirely.

In the standard (non-async) path, sglang_rollout.py::abort() already enforces this contract:

# sglang_rollout.py:376
if not args.partial_rollout:
    continue  # ABORTED samples are discarded, not collected

_AsyncRolloutWorker._make_done_cb does not apply the same check, so fully-async effectively forces partial_rollout=True behavior regardless of the user's configuration.

Root Cause

When an ABORTED group is requeued with existing sample.tokens (prompt + partial response), the next call to generate receives a non-empty sample.tokens and treats it as a prefix for continuation. With partial_rollout=False and mask_offpolicy_in_partial_rollout=False (the defaults), tokens generated under previous weight versions are appended without masking and participate in the loss.

Observed symptom — a single sample accumulating tokens across three weight versions:

weight_versions     = ["2", "4", "6"]
finish_reasons      = ["abort", "abort", "length"]
per-segment tokens  = [2, 2, 8]

Fix

Guard the requeue with the same partial_rollout flag used in abort():

if any(getattr(s, "status", None) == Sample.Status.ABORTED for s in result):
    if self.args.partial_rollout:  # align with sglang_rollout.py::abort()
        try:
            self.data_buffer.add_samples([result])
        except Exception:  # noqa: BLE001
            logger.exception("fully-async: failed to requeue aborted group")
    return

Behavior after fix:

• partial_rollout=False (default): ABORTED groups are dropped; the worker tops up from the buffer with a fresh sample, matching standard-path semantics.
• partial_rollout=True: unchanged — ABORTED groups are still requeued for continuation.

Verification

Run any fully-async training job with partial_rollout unset (default False) and frequent enough weight updates to produce ABORTED samples. Inspect the sample dump:

• Before fix: samples show multiple entries in weight_versions and finish_reasons starting with "abort"
• After fix: every completed sample has a single weight_versions entry