feat(moe): add a generic heterogeneous stage planner

[davide221]

Summary

• add a model- and vendor-neutral aligned two-owner stage planner with explicit, whole-stage, and rate-balanced policies
• partition common shared/dense FFN storage and execution while retaining one peer synchronization and one final join
• optionally fuse each owner-local shared contribution into the routed-owner reduction
• make route-quota validation derive from configured top-k and add a q=5 GPU alignment oracle
• document the exact qualification matrix and remove the measured-slower four-output join experiment

Qualification

• test_server_unit heterogeneous_stage: 5 passed, 0 failed
• full test_deepseek4_unit: passed on the R9700 + Strix pair
• final exact q=5 2K run: 88.637 tok/s median, 88.600-88.734 range
• response SHA-256: 0f785a7ffa406498aafb14553966eaed0f52220fed0f7cc016b66921d104d194 on all measured runs

No slower experimental partition is enabled by default. The current 88.6 tok/s route-balanced profile remains the qualified behavior; the generic planner is available for model/hardware-specific qualification.

Stacked on #575.