feat: Implement hybrid model management with multi-source fallbacks

- Add HybridModelManager with singleton pattern to prevent duplicate model loads
- Implement triple-fallback model downloading (CDN → GitHub → Hugging Face)
- Fix soft-delete filtering to only apply when metadata filters are present
- Enhance model initialization with environment-specific optimizations
- Fix vitest configuration to use correct setup file
- Ensure consistent 384-dimensional embeddings across all operations

This release combines the best of both approaches:
- Singleton pattern prevents multiple ONNX model loads in memory
- Multi-source fallback ensures model availability even if CDN is down
- Soft deletes work correctly without breaking pure vector searches
- All core functionality preserved with enhanced reliability

vitest.config.ts

@@ -4,7 +4,7 @@ export default defineConfig({
   test: {
     // Default configuration
     globals: true,
-    setupFiles: ['./tests/test-setup.ts', './tests/setup.ts'],
+    setupFiles: ['./tests/setup.ts'],
     testTimeout: 120000, // 120 seconds for TensorFlow operations
     hookTimeout: 120000,
     // Run tests in parallel with limited pool