cli: attach stderr stream of external diff command to pager

It's messy if pager and child output are interleaved as the pager controls
the tty.

Windows code is untested. I think the underlying I/O behavior is similar, but
I don't have expertise.

cli/src/merge_tools.rs

@@ -447,6 +447,7 @@ pub fn generate_diff(
     let mut child = cmd
         .stdin(Stdio::null())
         .stdout(Stdio::piped())
+        .stderr(ui.stderr_for_child().map_err(ExternalToolError::Io)?)
         .spawn()
         .map_err(|source| ExternalToolError::FailedToExecute {
             tool_binary: tool.program.clone(),

cli/src/ui.rs

@@ -186,6 +186,14 @@ impl Ui {
         }
     }
 
+    /// Stderr stream to be attached to a child process.
+    pub fn stderr_for_child(&self) -> io::Result<Stdio> {
+        match &self.output {
+            UiOutput::Terminal { .. } => Ok(Stdio::inherit()),
+            UiOutput::Paged { child_stdin, .. } => Ok(duplicate_child_stdin(child_stdin)?.into()),
+        }
+    }
+
     /// Whether continuous feedback should be displayed for long-running
     /// operations
     pub fn use_progress_indicator(&self) -> bool {
@@ -358,6 +366,18 @@ impl Drop for OutputGuard {
     }
 }
 
+#[cfg(unix)]
+fn duplicate_child_stdin(stdin: &ChildStdin) -> io::Result<std::os::fd::OwnedFd> {
+    use std::os::fd::AsFd as _;
+    stdin.as_fd().try_clone_to_owned()
+}
+
+#[cfg(windows)]
+fn duplicate_child_stdin(stdin: &ChildStdin) -> io::Result<std::os::windows::io::OwnedHandle> {
+    use std::os::windows::io::AsHandle as _;
+    stdin.as_handle().try_clone_to_owned()
+}
+
 fn term_width() -> Option<u16> {
     if let Some(cols) = env::var("COLUMNS").ok().and_then(|s| s.parse().ok()) {
         Some(cols)