Change init method from fork-server to spawn

Summary: It seems AMD GPU runtime doesn't quite work with forkserver for some reasons that we still need to debug. Basically hipMalloc will fail for subprocess. So just change it to spawn - I think it should be fine?

Differential Revision: D63311340

torchrec/distributed/test_utils/multi_process.py

@@ -126,7 +126,7 @@ def _run_multi_process_test(
         # pyre-ignore
         **kwargs,
     ) -> None:
-        ctx = multiprocessing.get_context("forkserver")
+        ctx = multiprocessing.get_context("spawn")
         processes = []
         for rank in range(world_size):
             kwargs["rank"] = rank