Minor performance improvements

[emlautarom1]

Summary

Small repeated-work items, collected so none is lost.

• Health selectors compile a Regex per (label × contain) per series per scrape: labels_contain calls Regex::new inside its double loop, invoked per metric series by count_labels/sum_labels, and the selectors themselves are rebuilt per check in health/checks.rs. Precompile or cache.
• CLI test commands build a fresh reqwest::Client per request — a new connection pool each time, so no keep-alive across RTT probes, which skews the very latency being measured: request_rtt (~20 call sites across beacon/mev/peers), relay_ping_test, and beacon.rs's http_client() per call site. request_rtt and relay_ping_test also carry no timeout (bounded only by cancellation) — one shared client with the 10s timeout the other paths already use fixes both. Improve alpha test peers: match Charon's probes; reduce timeouts #632 should adopt it for peers.
• DutyType's Display round-trips through serde_json per call (types.rs#L55-L65) and feeds ~25 metric-label call sites (scheduler, tracker reporters, bcast/consensus metrics, dutydb). Give it an as_str() like ProposalType already has.
• parsigex do_recv clones duty + par_sig per entry per inbound message (handler.rs#L264-L267) because the Verifier alias takes all three by value; narrowing to borrows ripples through every implementor and the boxed-future lifetime — decide once and do it.
• conn_logger still tests connectivity via the allocating connections_to_peer(..).is_empty() (conn_logger.rs#L266) — cold path, but PeerStore::has_connection exists for exactly this.