fix(gkd): preserve student gradients for empty token partitions

[Excelius-Wang]

When a context-parallel partition has no active response tokens, jsd_loss returns a detached zero. cp_reduce can give that rank a nonzero reduced loss, but backward still fails because the local student graph is disconnected. The same empty path is reached when all teacher top-k rows are uncovered.

Return the sum of the empty student logits so the loss remains zero and backward produces zero student gradients. Nonempty-token computation is unchanged.

Validation: CPU regression tests cover all three beta branches, full-vocabulary and top-k masking, uncovered top-k rows, and backward through a tiny linear model. The expanded suite covers FP16/BF16/FP32/FP64, vocabulary-size alignment, no-grad evaluation, and two AdamW updates with an empty accumulation microbatch. Restoring baseline jsd_loss gives 23 errors; all five test methods pass with the fix. Nonempty loss and gradient comparisons also pass. A two-process Gloo probe starts with a globally valid prompt/response label sequence and executes the production label shift, CP slicing and Megatron trainer loss method. The empty rank fails on the baseline and both ranks pass after the fix. The probe uses a tiny linear model and CPU-adapted index allocation, not a full Megatron runtime. Applicable pre-commit hooks pass. No GPU/Megatron end-to-end run.