The existing code admitted the following interleaving between
thread-1, running the async raft log truncation, and thread-2
which is running a new durabilityAdvancedCallback.

thread-1: executes queued := t.mu.queuedDurabilityCB and
 sees queued is false
thread-2: sees t.mu.runningTruncation is true and sets
 t.mu.queuedDurabilityCB = true
thread-1: Sets t.mu.runningTruncation = false and returns

Now the queued callback will never run. This can happen in tests
that wait for truncation before doing the next truncation step,
because they will stop waiting once the truncation is observed
on a Replica, which happens before any of the steps listed above
for thread-1.

Fixes cockroachdb#77046

Release justification: Bug fix

Release note: None