BeamNode mutex serializes chain work; heavy STF blocks interval vs gossip

## Summary

`BeamNode` uses a single `std.Thread.Mutex` to serialize work between the libxev main thread (`onInterval`) and the libp2p callback path (`onGossip`, `onReqRespResponse`, and locked sections of `onReqRespRequest`). Gossip handlers take the lock for the entire `chain.onGossip` path.

The Rust libp2p runtime runs on a dedicated OS thread, but **consensus CPU (state transition, verification, fork-choice updates) runs synchronously inside Zig callbacks while holding that mutex**. There is no separate worker pool for compute-heavy chain steps.

## Why it matters

Under CPU pressure (e.g. hash-sig verification, large blocks, deep fork-choice work), long critical sections can:

- stall `onInterval` while gossip holds the lock (or the reverse), worsening wall-clock vs head lag;
- contribute to RPC/gossip backpressure and timeouts observed on multi-client devnets.

## References (code)

- `pkgs/node/src/node.zig`: comment on `mutex` (serializes libxev main thread vs libp2p worker); `onGossip` locks before `chain.onGossip`; `onInterval` locks around `chain.onInterval`.
- `pkgs/network/src/ethlibp2p.zig`: dedicated thread for Rust network (`createAndRunNetworkThread`).

Fork choice has its own `RwLock` / `Mutex`, but that is orthogonal to this coarse `BeamNode` serialization.

## Suggested directions (for discussion)

- Shorten critical sections (e.g. parse/validate under lock, defer heavy work to a bounded queue + worker thread(s) with clear ordering rules).
- Or document this as an intentional invariant and tune timeouts / queue sizes accordingly.

## Context

Observed when analyzing devnet behavior: network and slot progression compete for the same lock while verification is expensive.

Read the mutex comment and callback implementations in pkgs/node/src/node.zig, especially onInterval, onGossip, onReqRespResponse, and onReqRespRequest. Then inspect createAndRunNetworkThread in pkgs/network/src/ethlibp2p.zig; the issue needs an agreed direction before work can be considered done, either reducing the critical sections with ordering guarantees or documenting and tuning the existing invariant.

Written by the indexing model from the issue text.