[Perf] Trim Code2Wav CUDA Graph buckets for Qwen3-TTS single-GPU deploy (#3932)

Signed-off-by: Rein Yang <ruiruyang2@gmail.com>

vllm_omni/deploy/qwen3_tts.yaml

@@ -28,6 +28,11 @@ connectors:
       codec_left_context_frames: 72
       # Emit only the first audio chunk early, then return to codec_chunk_frames.
       initial_codec_chunk_frames: 1
+      # Common Stage1 decode buckets:
+          #   no-ref first/steady chunks: 25 / 97 frames
+          #   Base ref-context first/steady chunks: 73 / 169 frames
+          #   decoder internal non-streaming chunks: 325 frames
+      decode_cudagraph_capture_sizes: [25, 73, 97, 169, 325]
 
 stages:
   - stage_id: 0