ShapeStream lifecycle tracked via scattered booleans instead of state machine

[KyleAMathews]

Summary

ShapeStream's lifecycle is tracked through independent boolean flags (#started, #connected) plus implicit state from #pauseLock.isPaused and #subscribers.size. Because these are independent, invalid state combinations are representable and can cause resource leaks.

Example: unsubscribeAll doesn't prevent restart

After unsubscribeAll():

1. Subscribers are cleared, timers are torn down
2. But #started remains true
3. If a pending pause lock releases, onReleased checks #started → true → calls #start()
4. #start() re-arms wake detection timers and makes HTTP requests with zero subscribers

Related: empty .catch(() => {}) blocks

Two places swallow all errors from #start():

• onReleased callback (line ~642): .catch(() => {}) with a comment claiming errors are "handled internally via onError" — but only if the user configured an onError handler. Without one, the stream silently freezes.
• requestSnapshot (line ~1744): same pattern.

These are pre-existing but become more impactful as more code paths flow through #start().

Potential approaches

Minimal fix: Add this.#started = false to unsubscribeAll() to prevent ghost restarts.

Deeper fix: Collapse the boolean flags into a state enum (idle | running | paused | stopped) so invalid combinations are unrepresentable. This would also make the empty .catch blocks easier to reason about — a stopped state would prevent restart attempts rather than relying on flag checks.

Context

Found during review of #3814 (wake detection) fix for pause/resume re-arming. The wake detection change slightly amplifies this issue since timers now re-arm in #start(), but the root cause predates it.