improve local scheduler log ordering (#366)

Summary:
Pull Request resolved: #366

This sets `PYTHONUNBUFFERED` to improve local scheduler logging latency. With this set each python log line gets immediately written out so you avoid large blocks of logs from the buffering. It also preserves groups unlike `readline()` since a stack trace will be written out as a single write

There's can still be some small amount of ordering mismatches but it's better than it was before

Reviewed By: kiukchung

Differential Revision: D33433782

fbshipit-source-id: 3f4a573a9736f1432119eeb9ad4e51e3bff19ada

Author: d4l3k

torchx/schedulers/local_scheduler.py

@@ -676,6 +676,9 @@ def _popen(
             else:
                 env["PATH"] = join_PATH(env.get("PATH"), cwd)
 
+        # default to unbuffered python for faster responsiveness locally
+        env.setdefault("PYTHONUNBUFFERED", "x")
+
         args_pfmt = pprint.pformat(asdict(replica_params), indent=2, width=80)
         log.debug(f"Running {role_name} (replica {replica_id}):\n {args_pfmt}")