Fix data race in async_scope

[lixin-wei]

This PR fixes a subtle data race in async_scope I introduced in a61207c.

The race occurs between __nest_rcvr::__complete() and __when_empty_op::start() (the on_empty() operation).

The Problematic Code Flow

Thread T2 (__complete)           Thread T1 (on_empty)
─────────────────────           ───────────────────
fetch_sub: 1→0
                                lock mutex
                                load __active (sees 0!)
                                unlock
                                start continuation → destroy scope
lock mutex → USE-AFTER-FREE

Solution

added another __pending_notifiers_ counter to guard this case.

I've tested this in my heavy workflow with TSAN.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ericniebler]

/ok to test e8d9f20