perf: optimize Docker build time from >10min to ~2min

- Remove sentence-transformers dependency (saves ~3GB PyTorch/CUDA)
- Make CrossEncoder import optional with graceful fallback
- Optimize Dockerfile for layer caching (incremental builds ~3s)
- Change PostgREST port from 3000 to 3001 (avoid Next.js conflict)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

config.toml

@@ -1,6 +1,6 @@
 model_provider = "litellm"
 # model = "anthropic/claude-sonnet-4-5-20250929"
-model = "hosted_vllm/openai/qwen-2.5-7b"
+model = "hosted_vllm/meta-llama/Llama-3.1-8B-Instruct"
 
 [projects."/home/valknar"]
 trust_level = "trusted"