Add support for sendmsg(2), recvmsg(2), and cmsg(3), take 2

[geofft]

I've confirmed that the following changes fix the reported Valgrind problem with PR #179, on both Linux and OS X 10.8.5 amd64:

@@ -122,3 +122,3 @@ struct cmsghdr {
     pub cmsg_type: c_int,
-    cmsg_data: [size_t; 0]
+    cmsg_data: [type_of_cmsg_len; 0]
 }
@@ -286,4 +286,6 @@ impl<'a> ControlMessage<'a> {
 pub fn sendmsg<'a>(fd: RawFd, iov: &[IoVec<&'a [u8]>], cmsgs: &[ControlMessage<'a>], flags: SockMessageFlags, addr: Option<&'a SockAddr>) -> Result<usize> {
+    let mut len = 0;
     let mut capacity = 0;
     for cmsg in cmsgs {
+        len += cmsg.len();
         capacity += cmsg.space();
@@ -317,3 +319,3 @@ pub fn sendmsg<'a>(fd: RawFd, iov: &[IoVec<&'a [u8]>], cmsgs: &[ControlMessage<'
         msg_control: cmsg_buffer.as_ptr() as *const c_void,
-        msg_controllen: cmsg_buffer.len() as size_t,
+        msg_controllen: len as size_t,
         msg_flags: 0,

Or in other words, fix the alignment of CMSG_DATA, and set msg_controllen to the actual length, not just the (rounded-up) length of the allocated buffer.

This is a resend of the reverted commit with those changes folded in (and, uh, fixing cmsg(4) to cmsg(3) in the commit message, idk why I thought it was in section 4).