fix: model weights setup — export scripts, ORT compat, bootstrap

- Add export_models.py for OpenCLIP ViT-B/32 and YOLOv8n ONNX export
- Fix ArgMax(13) ORT ARM64 incompatibility by passing eot_indices as a
  separate ONNX input (computed outside the graph in embed.py)
- Use legacy TorchScript exporter (dynamo=False) for IR version 9 compat
- Upgrade onnxruntime to 1.18.1
- Rewrite bootstrap_models.py with clear separation of auto-downloadable
  models (YuNet, SFace) vs manually-exported ones (OpenCLIP, YOLOv8n)
- Wire bootstrap into worker CMD (runs before Celery)

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

backend/app/services/vision/embed.py

@@ -71,8 +71,13 @@ class OpenCLIPEmbedder(Embedder):
         import open_clip
         tokenizer = open_clip.get_tokenizer("ViT-B-32")
         tokens = tokenizer([text]).numpy().astype(np.int64)
-        input_name = self._textual.get_inputs()[0].name
-        out = self._textual.run(None, {input_name: tokens})[0][0]
+        # Compute EOT indices outside ONNX (avoids ArgMax(13) op)
+        eot_indices = tokens.argmax(axis=-1).astype(np.int64)
+        inputs = self._textual.get_inputs()
+        out = self._textual.run(None, {
+            inputs[0].name: tokens,
+            inputs[1].name: eot_indices,
+        })[0][0]
         out = out / np.linalg.norm(out)
         return out.astype(np.float32)