transformers

b29c5c17 - Move the remaining router-masked models onto `ep_dispatch_experts`

Move the remaining router-masked models onto `ep_dispatch_experts` Gemma 4, DiffusionGemma, GraniteMoE-SWA, HY-V4, Inkling and MiMo-V2-Flash defaulted to router masking (`ep_router` + `moe_tp_experts`); their experts take expert ids like every other experts class, so they move to `ep_dispatch_experts`, which runs them masked when `ep_size == tp_size` and dispatches tokens otherwise. The router-masked plan stays for experts that route themselves (MXFP4 GPT-OSS). An `ep_router` rule next to `ep_dispatch_experts` is refused, since it would mask the tokens twice. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>