Add finish_from_stage4.sh: eval, bake, Hub upload without re-training

add scripts/run_resume_stage3.sh

resume script: fix dataset module names + opt-out of hf_transfer

add scripts/run_resume_stage2.sh

run_full_pipeline.sh: install hf_transfer + unsloth_zoo extras

run_full_pipeline.sh: pin trl/datasets/tyro to unsloth_zoo's window

training: push adapters to HF Hub after SFT + each GRPO stage

grpo: skip-SFT continuation script + completion-shape fix

run_full_pipeline.sh: pin transformers<5; drop xformers; +hf_hub

run_full_pipeline.sh: lock torch+torchvision; --no-deps for unsloth

run_full_pipeline.sh: install torchvision (unsloth import dep)

run_full_pipeline.sh: explicit unsloth_zoo install

run_full_pipeline.sh: fix legacy 'future' import error on Py3.10

run_full_pipeline.sh: add HF artifact upload step

OpenSOC v1