Fix path normalization for StreamWriter-based save operation (#3248)

Summary:
Follow up of https://github.com/pytorch/audio/issues/3243. Save compat module had different semantics than info and load, which requires different way of performing path normalization.

Pull Request resolved: https://github.com/pytorch/audio/pull/3248

Reviewed By: hwangjeff

Differential Revision: D44774997

Pulled By: mthrok

fbshipit-source-id: 4b967ae3ca6b45850d455b8e95aaa31762c5457e

torchaudio/_backend/utils.py

@@ -122,7 +122,7 @@ class FFmpegBackend(Backend):
         buffer_size: int = 4096,
     ) -> None:
         save_audio(
-            os.path.normpath(uri),
+            uri,
             src,
             sample_rate,
             channels_first,

torchaudio/io/_compat.py

@@ -204,8 +204,11 @@ def save_audio(
     bits_per_sample: Optional[int] = None,
     buffer_size: int = 4096,
 ) -> None:
-    if hasattr(uri, "write") and format is None:
-        raise RuntimeError("'format' is required when saving to file object.")
+    if hasattr(uri, "write"):
+        if format is None:
+            raise RuntimeError("'format' is required when saving to file object.")
+    else:
+        uri = os.path.normpath(uri)
     s = StreamWriter(uri, format=format, buffer_size=buffer_size)
     if format is None:
         tokens = str(uri).split(".")