【CI】add qwen3.5 cases

[kkscilife]

1. add qwen3.5 cases
2. fix gptoss config
3. TODO: update threshold to 1e-6

[kkscilife]

@claude review

[claude]

Claude: ## Summary
Adds two new CI test cases (qwen3-sft-cache, qwen3-5-sft-vl-moe), fixes gptoss config to use OpenaiTokenizeFunctionConfig, and adds image_pull_policy="Always" to cluster runner.

Issues

Warning

• autotest/config/qwen3_sft_cache.py:47 — Uses deprecated dataset_cfg parameter on TrainerConfig. Should use DataloaderConfig(dataset_config_list=...) instead (as done correctly in qwen3_5_30B_sft_vl.py).
• autotest/config.yaml:320-321 — grad_norm and loss/reduced_llm_loss thresholds for qwen3-5-sft-vl-moe are set to 5, effectively disabling validation. PR description notes "TODO: update threshold to 1e-6" — should this be resolved before merge?
• autotest/cluster/clusterx.py:53 — image_pull_policy="Always" is hardcoded for all jobs. Consider making it configurable via resource config.

Verdict

COMMENT — The core changes look correct. The deprecated API usage and loose thresholds should be addressed.