Restore GPT-OSS's router-masked expert-parallel plan
The MXFP4 experts route their tokens themselves: they take the routing data `routing_torch_dist` builds (already
restricted to the rank's experts) instead of expert ids, which token dispatch cannot exchange. With the plan on
`ep_dispatch_experts`, MXFP4 GPT-OSS under expert parallelism failed on its first forward
(`ep_forward() got an unexpected keyword argument 'scatter_idx'`). The router-masked plan from before #49160 leaves
them to their own routing.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
