fix: yield GIL in realtime idle wait — busy-wait starves co-resident threads

The realtime loop's idle branch (<|silence|>, nothing pending) spins on
`while True: continue`, monopolizing the CPython GIL. Every other thread in
the process is starved for as long as the session idles — e.g. an in-process
ASR decode inflates from ~150ms to 40-80s.

Measured (one busy-spin thread + SenseVoice decode in the same process):
584ms -> 42.6s (73x). Replacing the bare `continue` with a 20ms sleep poll
restores 132-152ms; input-detection latency stays negligible.

Platform-neutral: GIL behavior is identical on CUDA deployments — any
co-process embedding (ASR, TTS, metrics) is affected the same way.

modeling_moss_vl.py

@@ -3099,8 +3099,11 @@ class MossVLForConditionalGeneration(MossVLPreTrainedModel, GenerationMixin):
                     break
 
             if self.continue_generating and should_wait_for_new_input and not frames_to_process and not prompts_to_process:
-                # Busy-wait — matches VideoMllama reference. Caller controls cadence via
-                # `max_tokens_per_turn` sleeping in `_real_time_sample`.
+                # GIL-friendly wait: a bare `continue` busy-spin monopolizes the GIL
+                # and starves in-process siblings (in-process ASR decode inflates
+                # ~150ms -> 40-80s while a realtime session idles in <|silence|>).
+                # 20ms poll keeps input-detection latency negligible.
+                time.sleep(0.02)
                 continue
             break