cuda: sleep 30s instead of poll-fail; let model load handle lazy init

harden cuda wait: 180s timeout + torch.cuda.init() force

fix bf16 NaN: fp32 LoRA cast + grad checkpointing across SFT/DPO/GRPO; lower lora_r=32; adam_eps=1e-7; fix GRPO reward scenario routing

fix SFT NaN: max_grad_norm=1, lr=5e-5, lora_alpha=128 default; harden workdir

cuda init wait retry

clone repo first, push helper, fix data path

bold pipeline SFT+DPO+GRPO+eval, smoke flag for Colab test