fix(loss): preserve gradients for skipped listwise reranker batches

[Excelius-Wang]

Listwise reranker loss returns a new, unrelated zero tensor when no query group is retained. With the documented LISTWISE_RERANKER_MIN_GROUP_SIZE filter, one DDP rank can skip its groups while another still contributes a loss. The skipped rank produces no model gradients, leaving its peer waiting in backward.

Return a sum over an empty logits slice in both zero-loss branches. This preserves the autograd connection and supplies zero gradients without reading skipped scores. Retained query losses and averaging are unchanged.

Validation: four CPU tests pass, covering both zero branches, three floating dtypes, skipped nonfinite scores, retained loss/gradient equivalence and real two-process DDP with skipped and unequal-rank steps plus no_sync gradient accumulation. A separate time-bounded Gloo probe reproduces the upstream backward timeout and verifies identical gradients after the fix. Applicable pre-commit checks pass. No full pretrained-model or Megatron/NCCL training run was performed.

Additional verification uses 600 unchanged public ranking records from SciDocs, StackOverflow and AskUbuntu through the existing preprocessor and reranker grouping/collation methods, with a tiny deterministic text-feature scorer. Across three minimum-size settings, 1,762 retained cases match upstream loss and gradients exactly; 38 skipped cases now supply zero gradients. Original AskUbuntu records reproduce the asymmetric-rank backward timeout at the default threshold of 2; the fix completes successfully. These are grouping/autograd checks, not pretrained-model quality tests.

When an entire optimizer step has zero gradients, existing optimizer momentum or weight decay can still update parameters. This change restores connected-zero semantics; it does not add globally empty-step skipping. AdamW and momentum SGD match an explicit connected-zero reference.

A random one-layer BERT smoke test also reproduces the upstream timeout with find_unused_parameters both enabled and disabled. The fix passes both configurations and non-reentrant gradient checkpointing; all 25 parameter gradients match a cross-entropy reference. This is a CPU/Gloo model-structure check, not full Trainer.train or pretrained training.