Sample Curator v0.1.10

v0.1.10 — fix: load_model for worker-delegated ML kinds

CLAP, Whisper, Demucs, and HF LLM Reload/Retry now succeed in
production. The frozen sidecar can't load these libs locally — they
live in the runtime venv where the worker subprocess runs — so
_instantiate_model returns a marker for them and inference goes
through the worker.