feat: CLIP auto-GPU si >1.5 Go VRAM libre + index FAISS IVF 11.5x plus rapide

CLIP embedder : auto-détection GPU avec vérification VRAM disponible.
Si >1.5 Go libre → CUDA, sinon → CPU. Évite les OOM quand Ollama
utilise déjà la VRAM.

FAISS : migration Flat → IVF (116 clusters, nprobe=8).
Benchmark : 0.46ms → 0.04ms par recherche (11.5x).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

core/embedding/clip_embedder.py

@@ -58,9 +58,19 @@ class CLIPEmbedder(EmbedderBase):
                 "Install it with: pip install open-clip-torch"
             )
         
-        # Default to CPU to save GPU for vision models (Qwen3-VL, etc.)
         if device is None:
-            device = "cpu"
+            try:
+                import torch
+                if torch.cuda.is_available():
+                    free_vram = torch.cuda.mem_get_info()[0] / 1024**3
+                    if free_vram > 1.5:
+                        device = "cuda"
+                    else:
+                        device = "cpu"
+                else:
+                    device = "cpu"
+            except Exception:
+                device = "cpu"
         
         self.model_name = model_name
         self.pretrained = pretrained