Preserve tracing spans across spawns, instrument components, record missing metrics

Summary

Spans. The workspace has 60 production tokio::spawn/spawn_blocking sites and exactly 2 .instrument(...) calls (zero Span::current()/in_current_span). Every uninstrumented spawn drops the span context — which is also the mechanism behind #588's topic="" problem, since log-topic labels are span fields: spans and topics are the same fix. #[tracing::instrument] appears 33 times, 22 of them in one file (validatorapi/component.rs); scheduler, fetcher, bcast, tracker, dutydb, parsigdb, aggsigdb, consensus/qbft, dkg, parsigex, and priority have none. Highest-value uninstrumented spawns: the scheduler's per-slot detached duty callbacks (scheduler.rs#L173-L184), the actor loops (tracker, aggsigdb, prioritiser), and the duty-workflow wiring in wire.rs.

Defined-but-never-recorded metrics (grep-verified, zero recording sites):

• app_eth2_using_fallback — missing entirely, yet the health check that queries it already exists and can never fire. Lands with the eth2api multi-client(#611).
• p2p_peer_network_{sent,receive}_bytes_total — the BandwidthFactory mechanism exists but only the relay server wires it; Node::new (client path) has no bandwidth parameter (behaviour.rs#L98).
• p2p_peer_streams, p2p_relay_connections — declared in p2p/src/metrics.rs, recorded nowhere (Charon populates both in p2p.go).
• core_validatorapi_proxy_request_latency_seconds — the RAII recorder exists, no construction site.

Otherwise name-for-name parity with Charon's metric families is good; app_eth2_* request/error/latency coverage is opt-in per call site (~11 of 76 endpoints) and moves into the client wrapper.

Proposed change

1. Add a spawn helper attaching Span::current() (the #588 suggestion) and adopt it at the production spawn sites; set one root span per long-running component.
2. Add #[tracing::instrument] to the duty-pipeline entry points (fetch/propose/participate/broadcast/store paths), with skips to keep fields bounded.
3. Wire the four dead metric families above (or delete any that are deliberately out of scope, with a note).
4. While in layers/metrics.rs, cover the topic-resolution layer with tests: the TopicVisitor::record_debug fallback and nearest-enclosing-span resolution are untested (only same-span level filtering is).

Start with crates/core/src/scheduler.rs, wire.rs, the actor loops, and layers/metrics.rs to map spawn and span behavior; then inspect crates/p2p/src/metrics.rs, crates/app/node/behaviour.rs, and app health checks. Done means production spawns preserve context, the listed metric families are wired or explicitly removed, and topic-resolution tests cover the named fallbacks.