fix(finality-grandpa): eliminate data races in voter, timer, and test infrastructure

[dimartiro]

Changes

Fix all data races detected by go test -race ./pkg/finality-grandpa (137 warnings on development, 0 after this PR):

• voter.go — wakerChan.waker: changed from *waker to atomic.Pointer[waker] so the goroutine in start() and callers of setWaker() no longer race on pointer publication.
• voter.go — sharedVoteState.Get(): replaced ad-hoc svs.mtx (which never synchronized with the writer) with a snapshot pattern — the voter loop publishes a VoterStateReport into atomic.Pointer and Get() reads it without
  locking.
• voter.go — pruneBackgroundRounds: collect finalize notifications under inner.Mutex, release the lock, then invoke env.FinalizeBlock outside it. Holding the mutex across a user-supplied callback was both a deadlock hazard (a slow
  environment could stall observers and Stop()) and what forced Get() to give up on locking in the first place. A snapshot is also published right after releasing the lock so concurrent readers see fresh state even while we're inside a slow callback.
• voter.go — processBestRound: fixed a pre-existing missing inner.Unlock() on the error path from bestRound.poll.
• timer.go: expired is now atomic.Bool (was written under a mutex, read without). The one-shot close of wakerChan.in is now guarded by sync.Once, replacing the previous mutex + flag combo.
• environment_test.go — BroadcastNetwork: protected senders/history with a mutex, added a stop channel and a separate WaitGroup so forwarder goroutines exit before Stop() closes the receiver (previously caused send-on-closed-channel races and potential panics).
• voter_test.go — TestBuffered: the shared run bool is now atomic.Bool.

Tests

go test -race ./pkg/finality-grandpa

Should report ok with no WARNING: DATA RACE. Before this PR the same command produced 137 warnings.

Issues

ChainSafe/go-jam#523

[timwu20]

Suggest dropping this whole snapshot mechanism — publishVoterStateSnapshot, the defer in poll(), the voterStateSnapshot atomic.Pointer field, the buildVoterStateReport helper, and the lock-on-init in VoterState().

The Get() race is real, but the fix is one-line: have sharedVoteState reference v.inner (the writer's actual mutex — the old svs.mtx was a separate mutex that never synchronized with anything) and take it briefly in Get() while building the report.

Today the only callers of Get() are in voter_test.go. The one external consumer — SharedVoterState.reset(...) in internal/client/consensus/grandpa/grandpa.go — stores the handle but doesn't expose a reader yet (there's a // TODO: telemetry). Per-poll snapshot work + an extra field + ~25 lines of plumbing seems like a lot to carry for just tests.

[timwu20]

Nice cleanup overall — the race fixes in timer.go, wakerChan, environment_test.go, and the missing Unlock in processBestRound all look right.

My one piece of feedback is on the VoterState snapshot change — left an inline comment on publishVoterStateSnapshot. tl;dr: the underlying Get() race is worth fixing, but the snapshot/atomic.Pointer pattern feels heavy given that the only current callers of Get() are tests.

Otherwise LGTM.

[dimartiro]

@timwu20 changes applied. Could you please take one more look?