fix(tracing-core): prevent potential deadlock by retaining Dispatch instances until after lock is released

[u-kai]

Motivation

Dropping Dispatch instances during a call to register_callsite can result in a deadlock.

This occurs when a Dispatch's associated Subscriber implementation performs logging (e.g., via info!) in its Drop implementation.
Since info! internally triggers register_callsite, this can re-enter the same lock (LOCKED_DISPATCHERS) already held during interest rebuilding, causing a deadlock.

This PR addresses Issue #3269, which provides a minimal example of the issue.

Solution

This PR introduces a temporary DispatchGuard (a Vec<Dispatch>) to retain strong references to Dispatch instances during the for_each traversal used in interest rebuilding.

By ensuring that these Dispatch instances are dropped only after the write lock on LOCKED_DISPATCHERS has been released,
we prevent re-entrant locking in the Drop path of a Subscriber, avoiding the deadlock.