libbinder: fix RPC setup races

When setting up connections, there are a few cases where we take the
server lock and then we take the session lock. However, when a session
is shutting down, there is one case where we took the session lock and
then the server lock. This is a big no-no, and it was causing a
deadlock in the 'Fds' test (this creates many threads - but it is very
shortlived, the threads are still being setup on the server when the
process shutsdown, hitting the deadlock occassionally).

The solution to this involves keeping a little bit of extra state inside
of RpcSession directly to understand when it's shutting down. Also, we
now fully cleanup sessions before removing them during the shutdown
process.

From this point on, we should always take the server lock and then the
session lock in order to avoid races (never the session and then the
server).

Bug: N/A
Test: binderRpcTest
Test: while $ANDROID_BUILD_TOP/out/host/linux-x86/nativetest/binderRpcTest/binderRpcTest --gtest_filter="*Fd*"; do : ; done
Change-Id: I9144c43939c0640a2ec53f93f6e685ddce4b3e83