Track expert usage stats for MoE models across the mesh

[michaelneale]

What

For MoE models (e.g. Mixtral/Qwen-MoE/gpt-oss class), track which experts get activated during inference and surface that as a mesh-wide statistic — purely as a curiosity/observability feature.

Why

• MoE routing is invisible today. We split MoE models across nodes but have no idea how balanced the expert activation actually is.
• Interesting signal for split planning: if a handful of experts dominate, expert-aware placement could beat naive layer splits.
• Fun/insightful thing to show in the console — a live histogram of expert activation across the mesh.

Sketch

• Expose per-token expert selection counts from the embedded staged runtime (skippy) as a lightweight counter, off by default or cheap enough to always run.
• Aggregate locally per model, then gossip a compact summary (additive field, ignorable by older nodes) so any node can render mesh-wide totals.
• Surface via /api/status (or a dedicated stats route) and a small console visualization; optionally emit as OTLP metrics through skippy-metrics.

Notes

• Must be additive on the wire — older peers should ignore the new gossip field.
• Keep the hot path cost near zero; prefer counters incremented in the runtime over per-token event streams.
• Privacy: counts only, no prompt or token content.

[ndizazzo]

This makes sense - I think once we get #1134 through #1137 in, and #1167, we will have a really good foundation to add these kind of count metrics simply.