send mqtt disconnect on peer shutdown, derive keepalive from lease, add optional presence to mqp2p

[fabracht]

• Peer::shutdown sends MQTT DISCONNECT after deregistering, and dropping a Peer inside a Tokio runtime does too. Before, the mqtt5 client outlived the peer and kept reconnecting under mqp2p-<name>, so a peer that went offline and back online fought its old client for the session.
• PeerConfig::with_keepalive_secs, defaulting to a third of the lease (5 s for the default 15 s lease) instead of mqtt5's 60 s. The broker drops a silent client after 1.5 keepalives, checked once per keepalive, so a dead link is noticed in 7.5 to 12.5 s instead of the ~115 s measured in the c2-fusion harness.
• PeerConfig::with_presence (opt-in): follow MQDB's $DB/_presence/# and skip peers whose MQTT client has disconnected until they reconnect. Out-of-order events are ignored by ts, only mqp2p- client ids are tracked, and the table is capped at 4096 entries. The lease remains the backstop.
• README: liveness section and the ACL grants peers need (readwrite on mqp2p/responses/#, since MQDB 0.8.33 answers only on topics the requester may publish to).

Tests: shutdown and drop each send a clean DISCONNECT (presence event with unexpected: false), and presence hides a departed peer within 5 s despite a 60 s lease and shows it again on return. All three fail without the disconnect; the presence test also fails without with_presence. Unit tests cover presence ordering, the cap, the client-id filter and the keepalive default. Workspace: 240 passed, 0 failed. The mqp2p integration tests need an installed mqdb binary and are not run in CI; they passed 5 times in a row locally against mqdb 0.8.51.