ci: use pushed tag for goreleaser and add manual dispatch

ci: poll for go-algorand releases every 10 minutes

v1.7

PR #1075 is a proper fix, but this will unblock PR #1074

innerStop and DisconnectPeers drop peers from the peer list directly.
Each peer's own removePeer then finds it gone and releases nothing, so
its identity and throttle slot leaked. Both networks now release a
peer through unregisterPeerLocked, from removePeer and from the
sweeps.

unregisterPeerLocked also marks the peer closing.
identityVerificationHandler checks that mark under the peer lock, so a
verification message queued before the peer left cannot claim the
identity again.

P2PNetwork.innerStop sweeps wsPeersToIDs rather than wsPeers. A wsPeer
replaced by a newer stream from the same peer is only in wsPeersToIDs,
so the old sweep left it open with its slot.

Stop sweeps the peer list before it shuts down the ws listener or the
p2p host. A handshake that completes in between registers a peer that
nothing closes, so the peer outlives Stop.

innerStop now sets a flag under the peer lock, and addPeer rejects
peers once it is set. The caller closes a rejected peer and skips the
connected logs and telemetry.

Found while reviewing #6744.

The read and write loops report errors only while didSignalClose is
unset. A swept p2p peer's streams reset before closeWaiter reaches
Close, which used to log a read error and count a dropped connection
on every shutdown.

The ws sweeps also clear the peer slice, so it keeps no swept peers
reachable, and update the connection gauges.

A peer's read and write loops started in init, before addPeer
registered it. tryConnect and the p2p stream handler also claimed the
identity in a separate critical section from the insert. A peer could
therefore receive messages, or close itself, before it was registered,
and a rejected peer needed cleanup in several places.

addPeer now calls init itself. It first skips a peer whose identity
another peer already holds, so a duplicate allocates no send queues.
Then, under the peer lock, it claims the identity and the throttle
slot, registers the peer and starts its loops. It reports whether it
admitted the peer, and the caller closes the connection of a peer
rejected as a duplicate. P2PNetwork gets the same addPeer.

Nothing can close a peer before addPeer now, so the closing check in
addPeer and algod_network_peer_already_closed go. So does the
duplicate pointer check, which could never match.

Tests that registered bare peers now build them with makeTestPeer, or
no longer register peers they only use as message senders.
TestPeeringWithIdentityChallenge expects one fewer setIdentity call,
because the duplicate no longer reaches it.

This prevents the hardcoded lastblock value causing an error when the
test data doesn't go to that round