Fix SIGPIPEs associated with parts

[mxw]

Parting sometimes causes SIGPIPEs, e.g., when the write&flush within g_io_channel_shutdown is invoked via paxos_drop_connection.

[mxw]

They also occur independently of g_io_channel_shutdown; I just ran traces/1.trace; when screen 2 parted, screen 1 got a SIGPIPE when it tried to write. I'm guessing that maybe the destruction of peers isn't total? Or maybe we try to write to a dead peer before its death is noticed and paxos_drop_connection is invoked by the I/O functions?

[zenazn]

I'm pretty sure SIGPIPEs only occur for unix domain sockets, and won't be an issue when we switch to TCP/UDP/???, which don't error out through signals.