fix(network): expire stale and unreachable peers from the network participant list

[lukyrys]

The network participant list (Settings → LISH network → Participants) kept showing peers that went offline long ago, and the backend kept re-dialing them forever. This makes those entries expire, and fixes the surrounding lifecycle and cleanup paths the work uncovered.

• a discovered peer is aged from its last successful contact, not from its last update, so repeated gossip and its own failed dials no longer keep a dead row alive; rows expire 30 minutes after that, scoped per network, and are capped at 256 per network
• a peer is only evicted from the peer store on real dial failures against addresses this host could route to — losing our own route no longer counts as the peer being dead — and an evicted peer is briefly quarantined so gossip cannot instantly re-add it
• configured bootstrap peers are never evicted, and only a dial-verified identity gets into the autodial list or exempts a row from expiry
• leaving a network now finishes even when one peer teardown fails, drops that network's bootstrap addresses, keeps peers shared with another network connected, and no longer reports participants for a network this node is not in
• network start/stop and enable/disable are serialised and bound to the run they began in, a failed stop is reported instead of being treated as a stopped node, and pending work is drained before shutdown
• peer announce never advertises an address belonging to another identity, bounds the input before parsing, caps addresses per message, and rate limits intake per source
• a factory reset treats a missing prepare step as a failed barrier instead of a passed one, so a category is only wiped once the node is proven stopped; categories are independent and on-disk files are never touched
• one downloader throwing during a leave no longer aborts the downloaders after it, and a download sourced from another joined network keeps running
• self-addresses are recognised by their destination identity, so a relayed entry naming this node only as the relay hop is no longer discarded
• frontend: factory reset surfaces which step failed (three new translation keys), no other UI changes