FLARE.py

Run GPU smoke checks for lra and pdebench models in this repo (compile, instantiate, strict state_dict reload, and CUDA forward pass). Use this whenever validating that model wiring/loading/execution still works after refactors or syncs.