Consolidate instance lifecycle subscriptions

[sjmiller609]

Summary

• replace the old per-instance state subscription path with one shared lifecycle event bus in lib/instances
• migrate WaitForState to filter the lifecycle bus and keep its polling fallback
• add lifecycle bus backpressure metrics for subscriber count, max queue depth, and dropped events

Why This Change

We were about to introduce a second internal subscription mechanism for the upcoming auto-standby work. The existing WaitForState path used a per-instance state subscription that only carried State/StateError, while the new controller needs a global stream of lifecycle changes like create, update, fork, and delete.

Instead of keeping two different in-memory buses, this change consolidates on one richer lifecycle event stream and makes WaitForState a consumer of that shared path. That keeps the public API the same, reduces duplicated fanout logic, and gives us one place to observe subscriber backpressure.

Testing

• go test ./lib/instances -run 'TestWaitForState|TestLifecycleSubscribers|TestLifecycleEventMetrics'
• go test ./lib/builds -run 'Test'
• go test ./lib/instances -run '^$'
• go test ./lib/builds -run '^$'
• go test ./cmd/api/api -run '^$'

---

Note

Medium Risk
Changes core instance lifecycle signaling by replacing per-instance state subscriptions with a shared lifecycle event bus, which could affect wait/polling behavior and internal controllers if events are missed or dropped. Adds new buffering/config/metrics paths that need validation in production workloads.

Overview
Consolidates internal instance change notifications by replacing the per-instance StateChange subscription mechanism with a shared LifecycleEvent bus, and updates the Manager interface to expose SubscribeLifecycleEvents.

WaitForState now consumes this shared stream (filtering by instance_id and handling delete explicitly) while retaining the 5s polling fallback, and manager mutation methods (CreateInstance, UpdateInstance, Start/Stop, Standby/Restore, Fork*, DeleteInstance) publish lifecycle events on success.

Adds observability and tuning for backpressure: new metrics for subscriber count, max queue depth, and dropped events, plus a new instances.lifecycle_event_buffer_size config setting (default 256) wired through providers into instances.ManagerConfig and validated in config/tests.

^{Reviewed by Cursor Bugbot for commit 134628d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[hiroTamada]

solid refactor — clean fan-out bus with good observability. two minor comments.