fix: enable CUDA graph capture with DeepSpeed hybrid engine
- Remove DS forward hooks (tput_timer.synchronize()) during graph capture
- Fix trainer: sync_weights_from_student -> sync_weights, add eval/train toggle
- Wire use_graph_capture through OPSD config and rollout wrapper
- Add smoke_hybrid_gc.json + smoke_ds_zero0.json configs for testing
The root cause was DeepSpeedEngine._module_forward_pre_hook calling
synchronize() via tput_timer.start(), which is illegal during CUDA
graph capture. We temporarily remove hooks for the capture+decode
section and restore them in the finally block.
