Extract a shared routed-expert FFN helper and fix the fmt=6 input rotation

[lineape]

What to build

Eliminate the parallel FFN copy — the root of the maintainer's objection and the fmt=6 bug — by extracting the four-step routed-expert FFN (expert_gate_up → siluf → e8_rot_rows → matmul_qt) into ONE shared helper called by both moe()'s CPU path and cluster_worker_run, and applying the shared E8_XE input rotation on both sides. Do NOT patch the worker's copy in place; do NOT introduce a new parallel implementation ("two implementations of the same thing can't both land").

This fixes the missing fmt=6 input pre-rotation by construction: the worker feeds raw x to gate/up today, while moe() rotates the input first (E8_XE, guarded on g.fmt==6), then rotates the intermediate before the down matmul. Sharing one helper makes that drift impossible.

Acceptance criteria

• moe()'s CPU path and cluster_worker_run call the same FFN helper (no duplicate four-step sequence anywhere).
• fmt=6 (E8/IQ3) experts: worker output == local moe() output (the bug is gone).
• fmt=4 (int4) experts unchanged (control arm — no rotation, results identical).
• Per-layer input-rotation caching preserved (don't regress fmt=5: a deployable E8-lattice grouped container — int3 that beats shipped int4 quality at 25% fewer bytes (design + plan) JustVugg/colibri#452's ~11 ms-per-expert cost).
• Token-exact: existing oracle still ~30-32/32 TF + 20/20 greedy; make check green, 0 warnings.

Environment: codegraph-indexed — run codegraph explore first; refs in /home/eli/dev/_reference-docs-and-repos/colibri.