fix(advisory-lock): make TryAttainLockAsync idempotent against re-entrant calls

[jeremydmiller]

Summary

• Postgres session-level advisory locks stack (Postgres docs: "Multiple lock requests stack, so that if the same resource is locked three times it must then be unlocked three times to be released."). SQL Server session-scoped application locks have the same reentrancy semantics. The heartbeat lease-renewal change in a84d6a262 calls TryAttainLeadershipLockAsync on every tick — including ticks where the leader already holds the lock — to refresh the lease for lease-based backends (RavenDb, Cosmos). For advisory-lock backends (Postgres, SQL Server) this stacks the lock by one count per heartbeat. The single ReleaseLeadershipLockAsync call during DisableAgentsAsync / stepDownAsync then only decrements once, leaving the lock held server-side and silently blocking failover — no error logged, just a stalled election.
• Surfaced in SlowTests.SharedMemory.leadership_compliance.take_over_leader_ship_if_leader_becomes_stale, which fails 0/10 in isolation on main (consistent regression since a84d6a262, bisected) and passes 3/3 with this fix. The full Postgres LeaderElection compliance suite (12 tests) is green with this change.
• Both AdvisoryLock.TryAttainLockAsync (Postgres) and SqlServerAdvisoryLock.TryAttainLockAsync get the same idempotent short-circuit: if _locks already contains the id and HasLock confirms the held connection is still alive, return true without calling pg_try_advisory_lock / sp_getapplock again. The HasLock liveness ping keeps the Leader split-brain: stale Postgres advisory lock causes two leaders to assign agents concurrently #2602 split-brain detection intact — if the lock was lost server-side, HasLock clears _locks and the next TryAttainLockAsync will actually re-attain.

Test plan

• New Bug_advisory_lock_stacking_blocks_failover regression test for both backends — verified to fail on unpatched main (proves they exercise the bug) and pass with the fix
• Bug_split_brain_advisory_lock_state_divergence (Leader split-brain: stale Postgres advisory lock causes two leaders to assign agents concurrently #2602 regression) still passes — the HasLock liveness ping is still exercised by the next TryAttainLockAsync after a backend kill
• Bug_2518_concurrent_migration_advisory_lock still passes
• SlowTests.SharedMemory.leadership_compliance.take_over_leader_ship_if_leader_becomes_stale — 3/3 pass (was 0/10 on unpatched main)
• Full Postgres LeaderElection compliance suite — 12/12 passed, 0 failed (~2m13s)

🤖 Generated with Claude Code