fix(bookmark): Run async on_restore callbacks before the session's effects

[jat255]

Async bookmark on_restore callbacks must finish before the session's other effects and outputs run. With the flush from #2508, those effects started early and saw state that was not restored yet. This PR makes them wait again.

Refs #2508, for its TODO "Ordering for async bookmark on_restore (priority 1000000)". Stacked on #2522.

Summary

On main, the flush awaited each effect in turn. The on_restore effect has priority 1000000, so its callbacks finished (awaits included) before any other effect started. With #2508, the flush starts each effect in priority order but does not wait for its async part. When an on_restore callback awaits, the session's other effects start first. They run once with default values, and then again after the restore. shinychat's enable_bookmarking() registers callbacks that await, so chat apps get this problem.

This PR adds a gate for each session:

• The restore effect resolves a future when its callbacks finish, also when its run is cancelled (shiny/bookmark/_bookmark.py).
• Each effect of the same session with a lower priority waits for this future before its body runs (shiny/reactive/_reactives.py). The effects still start in priority order. They keep the session busy, so no outputs go out before the restore is complete.
• Session._restore_gate() connects the two, and SessionProxy sends module effects to the root session (shiny/session/_session.py).

Review notes

• The gate is per session, so a slow restore does not delay other sessions. I did not make the flush wait for the restore effect, because the flush is global and that would block all sessions.
• Effects with a priority of 1000000 or more do not wait. This matches main, where they ran before the restore effect.
• The gate uses the priority at which the run was queued. set_priority() takes effect on the next invalidation, so a changed priority cannot skip the gate.
• An alternative is to await the callbacks in the init handler before the first flush. I did not use it, because it blocks the session's receive loop during the restore.
• R has no equivalent, because its onRestore() callbacks are synchronous.
• There is no CHANGELOG entry. Compared with the released version, the behavior does not change.
• A duck-typed session mock now needs _restore_gate(), in the same way that it needs _otel_reactive_update_span(). The mock in tests/pytest/test_poll.py has one now.

Testing

• 10 new tests in tests/pytest/test_concurrency.py. They cover the first output and effect values, priority order, effects above priority 1000000, other sessions, a callback that raises, a cancelled callback, the session end during a restore, module effects, a waiting effect that is destroyed, and set_priority(). Without the fix, 9 fail. The other one guards the priority 1000000 exception. I also removed the SessionProxy delegation, the finally, and the asyncio.shield() one at a time, and each change made its test fail.
• The full pytest suite and the Playwright bookmark tests (chromium) pass. check-format, check-lint, and pyright are clean. pyrefly reports 0 errors.