libbinder: no temp rpc sess leak w spurious wakeup

RpcSession incoming threads continued to hold an RpcSession
instance after they set the shutdown condition (removing the
associated 1:1 thread connection object from RpcSession's
mConnections object). Since the shutdown condition must
include cleaning up RpcSession, we cannot delay or remove
clearing the associated connections. Instead, a new explicit
shutdown condition is added, which does not restrict the
manipulation of the session object.

Interestingly, this issue was blocking several changes on-and-off
for a few weeks. Though, test hub doesn't show it failing at all. I
couldn't reproduce it locally even with 5 days (24hr/day) and
one of these failing tests running in a tight loop, with builds
running in the background (devinmoore@ reported a local failure
with a build running). I submitted several changes to debug this,
and one of them (that dumped backtraces), should have caught it,
except the race is just too rare.

When we have this situation: a retrospectively benign problem causing
a big failure, the obvious question to ask is, is the test too
brittle? No! If this is the sensitivity at which it finds a bug, we
can hardly imagine an error going unnoticed here. Only if this
situation repeated several times or some of these issues became too
plenty to maintain would I think that we needed to "tone down the
tests".

Finally, how did this get fixed: dump every incStrong backtrace in
RpcSession and investigate all the code that is responsible for
maintaining those sp<>s. Wheeee :)

Bug: 244325464
Test: binderRpcTest

Change-Id: I76ac8f21900d6ce095a1acfb246ca7acf1591e0b