poll: Do not clear readiness on short read/writes.

The new mio_unsupported_force_poll_poll behaviour works the same as
Windows (using level-triggered APIs to mimic edge-triggered ones) and it
depends on intercepting an EAGAIN result to start polling the fd again.

Closes #5866

tokio/src/io/poll_evented.rs

@@ -166,7 +166,7 @@ feature! {
                     Ok(n) => {
                         // if we read a partially full buffer, this is sufficient on unix to show
                         // that the socket buffer has been drained
-                        if n > 0 && (!cfg!(windows) && n < len) {
+                        if n > 0 && (!cfg!(windows) && !cfg!(mio_unsupported_force_poll_poll) && n < len) {
                             self.registration.clear_readiness(evt);
                         }
 
@@ -197,7 +197,7 @@ feature! {
                     Ok(n) => {
                         // if we write only part of our buffer, this is sufficient on unix to show
                         // that the socket buffer is full
-                        if n > 0 && (!cfg!(windows) && n < buf.len()) {
+                        if n > 0 && (!cfg!(windows) && !cfg!(mio_unsupported_force_poll_poll) && n < buf.len()) {
                             self.registration.clear_readiness(evt);
                         }