[Edge][WP2] Add initialization timeline and cold-start profiler

* vllm_omni/engine/init_timeline.py: InitTimeline phase collector; engine
  phases (stages_init, g3_device_lock, g2_spawn/ready) and worker phases
  (proc_entry, engine_core_init, init_device, load_model/load_weights,
  talker_mtp_init, attn_prewarm, cudagraph_capture, talker_mtp_capture,
  code2wav_cudagraph, predictor_setup_compile/warmup/cudagraph) recorded via
  worker_phase/timed_phase; workers append JSONL under
  VLLM_OMNI_INIT_TIMELINE_DIR (inherited at spawn) and the engine merges them.
  Enabled by VLLM_OMNI_INIT_TIMELINE=<json> or log_stats; table logged at INFO,
  partial table on orchestrator startup timeout, suggested timeouts logged.
* omni_base.py: named deploy profiles default to stage_init_timeout=900 /
  init_timeout=1500 unless given explicitly.
* gpu_model_runner._init_talker_mtp tolerates max_cudagraph_capture_size=None;
  gpu_ar_model_runner.shutdown skips CUDA graph teardown on CPU.
* benchmarks/tts/init_profile.py: runs stream_latency_bench --init-only per
  configuration (default, cold_cache, eager_stage1, eager_both,
  parallel_stage_init, edge, ...) and tabulates phase durations.
* stream_latency_bench: --parallel-stage-init / --deploy-profile flags; the
  resource poller no longer shadows Thread._stop and skips GPU sampling on CPU.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

vllm_omni/model_executor/models/common/qwen3_code_predictor.py

@@ -26,6 +26,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmb
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 
 from vllm_omni.diffusion.layers.custom_op import CustomOp
+from vllm_omni.engine.init_timeline import timed_phase
 from vllm_omni.platforms import current_omni_platform
 
 logger = init_logger(__name__)
@@ -716,6 +717,7 @@ class CodePredictorWrapper(nn.Module):
             return
         self._proj_buf = torch.zeros(bsz, max_seq, self._cp_hidden, dtype=dtype, device=device)
 
+    @timed_phase("predictor_setup_compile")
     def _setup_compile(self) -> None:
         """Lazily set up torch.compile with optional device graph capture."""
         if self._compiled_model_fwd is not None:
@@ -842,6 +844,7 @@ class CodePredictorWrapper(nn.Module):
         allowed = set(all_seq_lens)
         return sorted(seq_len for seq_len in self._prefix_graph_seq_lens if seq_len in allowed)
 
+    @timed_phase("predictor_warmup")
     def _warmup_buckets(self) -> None:
         """Warmup power-of-2 batch-size buckets to front-load Inductor compilation."""
         max_bsz = self._vllm_config.scheduler_config.max_num_seqs
@@ -897,6 +900,7 @@ class CodePredictorWrapper(nn.Module):
                     self._compiled_model_fwd(proj_buf[:bsz, :max_seq, :], pos_ids)
             logger.info("code_predictor: warmup done for buckets %s", self._bucket_sizes)
 
+    @timed_phase("predictor_cudagraph")
     def _capture_cuda_graphs(self) -> None:
         """Capture a CUDA graph per bucket using vLLM's global graph pool."""
         from vllm.platforms import current_platform