firmware: fix TOCTTOU race in sync/async RPC code.

Before this commit, the main loop in session code was laid like:

  1. process_kern_queued_rpc
  2. process_host_message
  3. process_kern_message

If a host message (such as an RPC reply) caused the kernel to exit,
then any async RPCs would not complete, since RunFinished immediately
shuts down the kernel.

Fix this by reordering 1 and 2.

artiq/firmware/runtime/session.rs

@@ -497,16 +497,16 @@ fn host_kernel_worker(io: &Io,
     let mut session = Session::new(congress);
 
     loop {
-        while !rpc_queue::empty() {
-            process_kern_queued_rpc(stream, &mut session)?
-        }
-
         if stream.can_recv() {
             process_host_message(io, stream, &mut session)?
         } else if !stream.may_recv() {
             return Ok(())
         }
 
+        while !rpc_queue::empty() {
+            process_kern_queued_rpc(stream, &mut session)?
+        }
+
         if mailbox::receive() != 0 {
             process_kern_message(io, Some(stream), &mut session)?;
         }