[client] Take more advantage of pipelined NUSB sending

[jamesmunns]

Right now, the transmitting of data with NUSB generally awaits the transfer to be complete. We could instead just submit, and then return, allowing multiple submissions to be sent nose-to-tail.

I tried this diff, and it didn't seem to help as much as I would have hoped, but we might need #69 to realize the improvements. One downside of this is that if there is an error, it will show up at the "wrong time", though it's likely that the input worker will notice a disconnect first anyway.

 impl NusbWireTx {
     async fn send_inner(&mut self, data: Vec<u8>) -> Result<(), NusbWireTxError> {
-        self.boq.submit(data);
-
-        let send_res = self.boq.next_complete().await;
-        if let Err(e) = send_res.status {
-            tracing::error!("Output Queue Error: {e:?}");
-            return Err(e.into());
+        while self.boq.pending() != 0 {
+            let send_res = self.boq.next_complete().await;
+            if let Err(e) = send_res.status {
+                tracing::error!("Output Queue Error: {e:?}");
+                return Err(e.into());
+            }
         }
+        self.boq.submit(data);
         Ok(())
     }
 }

[jamesmunns]

Ah, this isn't good either, we still wait for pending frames to complete. That explains the lack of perf difference. We might want to select on completion and maybe another intermediate channel? The worker tasks being outside of the interfaces may have been the wrong place to draw the interface line.