Fix RPC write_message to use unbuffered binary I/O on Windows

The write_message function was using text mode (sys.stdout.write) which
causes Python to perform automatic line-ending translation on Windows.
This corrupts the JSON-RPC protocol headers which expect exact byte sequences
like 'Content-Length: N\r\n\r\n'.

Use unbuffered binary I/O (os.write) instead, mirroring the pattern already
used by read_message and read_message_with_timeout which use os.read().
This ensures:
- No platform-specific line-ending translation
- Symmetric read/write I/O patterns
- Atomic writes for small RPC messages (within PIPE_BUF)
- Single system call (same as before)

Fixes the 10-second timeout failures on Windows runners where the Python
RPC server couldn't send responses to Java due to protocol header corruption.

Author: knutwannheden

rewrite-python/rewrite/src/rewrite/rpc/server.py

@@ -1380,14 +1380,15 @@ def read_bytes(n: int) -> Optional[bytes]:
 
 
 def write_message(response: dict):
-    """Write a JSON-RPC message to stdout."""
-    content = json.dumps(response)
-    content_bytes = content.encode('utf-8')
-
-    sys.stdout.write(f"Content-Length: {len(content_bytes)}\r\n")
-    sys.stdout.write("\r\n")
-    sys.stdout.write(content)
-    sys.stdout.flush()
+    """Write a JSON-RPC message to stdout.
+
+    Uses unbuffered binary I/O to avoid line-ending translation on Windows
+    that would corrupt the JSON-RPC protocol headers. Mirrors the pattern
+    used by read_message() which uses os.read() on the read side.
+    """
+    content_bytes = json.dumps(response).encode('utf-8')
+    header = f"Content-Length: {len(content_bytes)}\r\n\r\n".encode('utf-8')
+    os.write(sys.stdout.fileno(), header + content_bytes)
 
 
 def main():