fix(sp): exclude internal ring padding from MoE auxiliary loss

[Excelius-Wang]

Ring sequence parallelism pads each packed sequence independently. The MoE output hook currently trims gathered router logits to the original total length, retaining internal padding and dropping valid tokens from later sequences. This changes the load-balancing loss and its gradients when router auxiliary loss is enabled.

Filter the gathered router logits using the padded position-ID mask in the ring path, reusing the existing padding layout. Keep the Ulysses-only trimming behavior.

Validation:

• CPU Gloo regression tests cover ring-only, ring + Ulysses, and Ulysses-only groups; compare router logits, Qwen3-MoE auxiliary loss, and local gradients against an unpadded reference across five packing layouts, including a single sequence; also cover disabled sequence parallelism. Before: two ring configurations fail; after: all four configurations pass.
• Related tests: 46 passed, 4 CUDA-only tests skipped.
• pre-commit run --all-files: passed.

No full GPU distributed-attention training run was performed.