fix: restore previous behavior on live pipeline + audio downscaler (#561)

This commit restore the original behavior with frame cutting. While
silero is used on our gpu for files, look like it's not working great on
the live pipeline. To be investigated, but at the moment, what we keep
is:

- refactored to extract the downscale for further processing in the
pipeline
- remove any downscale implementation from audio_chunker and audio_merge
- removed batching from audio_merge too for now

server/gpu/modal_deployments/reflector_transcriber_parakeet.py

@@ -172,7 +172,7 @@ class TranscriberParakeetLive:
         text = output.text.strip()
         words = [
             {
-                "word": word_info["word"],
+                "word": word_info["word"] + " ",
                 "start": round(word_info["start"], 2),
                 "end": round(word_info["end"], 2),
             }
@@ -213,7 +213,7 @@ class TranscriberParakeetLive:
 
             words = [
                 {
-                    "word": word_info["word"],
+                    "word": word_info["word"] + " ",
                     "start": round(word_info["start"], 2),
                     "end": round(word_info["end"], 2),
                 }
@@ -386,7 +386,7 @@ class TranscriberParakeetFile:
                 text = output.text.strip()
                 words = [
                     {
-                        "word": word_info["word"],
+                        "word": word_info["word"] + " ",
                         "start": round(
                             word_info["start"] + start_time + timestamp_offset, 2
                         ),