fix: standalone inference.py with raw websockets, correct obs/state parsing

fix: reconnect per task to avoid closed WebSocket on second episode

fix: zero-arg self-contained graders, inference runs all 3 tasks

fix: rename API_KEY to HF_TOKEN, remove fallbacks, fix action spaces in STEP log

Fix inference.py: rename IMAGE_NAME to LOCAL_IMAGE_NAME, move debug logs after [START]

Fix task selection: reset() accepts task param, all 3 tasks verified

Rebuild inference + client to match OpenEnv WebSocket spec

OpenEnv Submission

update inference

Initial commit