Add vision model enhancements and image processing capabilities

- Update `requirements.txt` to include Pillow for image handling.
- Refactor vision model validation logic in `voice_webrtc.py` to improve error handling for unsupported image input.
- Introduce new functions in `pipeline.py` for image data processing and analysis using vision models.
- Implement `VisionCaptureProcessor` to manage video frame requests for auxiliary vision model analysis.
- Enhance the pipeline to support image input requests and integrate vision model responses into the processing flow.

backend/routes/voice_webrtc.py

@@ -74,16 +74,6 @@ async def _resolve_config(offer: SignalingOffer) -> AssistantConfig:
     raise ValueError("offer 缺少 assistant_id 或 inline_config")
 
 
-def _apply_vision_model(cfg: AssistantConfig) -> None:
-    cfg.model = cfg.vision_model
-    cfg.llm_interface_type = cfg.vision_llm_interface_type
-    cfg.llm_values = cfg.vision_llm_values
-    cfg.llm_secrets = cfg.vision_llm_secrets
-    cfg.llm_support_image_input = cfg.vision_llm_support_image_input
-    cfg.llm_api_key = cfg.vision_llm_api_key
-    cfg.llm_base_url = cfg.vision_llm_base_url
-
-
 async def _handle_offer(websocket, payload, peers):
     from pipecat.transports.smallwebrtc.connection import SmallWebRTCConnection
     from services.pipecat.pipeline import run_pipeline
@@ -99,11 +89,17 @@ async def _handle_offer(websocket, payload, peers):
         cfg = await _resolve_config(offer)  # 解析放在建连前,配置错就别建连
         vision_enabled = offer.vision_enabled or cfg.vision_enabled
         if vision_enabled:
-            if not cfg.vision_llm_support_image_input:
+            has_native_vision = (
+                not cfg.vision_model_resource_id and cfg.llm_support_image_input
+            )
+            has_aux_vision_model = (
+                bool(cfg.vision_model_resource_id)
+                and cfg.vision_llm_support_image_input
+            )
+            if not (has_native_vision or has_aux_vision_model):
                 raise ValueError(
-                    "当前视觉模型不支持图片输入,请在模型资源中选择支持图片输入的 LLM"
+                    "当前模型不支持图片输入,请在模型资源中选择支持图片输入的视觉模型"
                 )
-            _apply_vision_model(cfg)
         pc = SmallWebRTCConnection(ice_servers=aiortc_ice_servers())
         if pc_id:
             pc._pc_id = pc_id