[script]: add Kimi LoRA script

[yyj6666667]

Summary

• Add a Kimi K2.5 LoRA tutorial covering data preparation, training, saving and inference loading.
• Prepare native non-thinking data once for unmodified LLaMA-Factory main.
• Update the two 8-GPU smoke YAMLs; packing remains disabled.

Only five example files change. No training runtime, dependency or CI changes.

Validation

• 13 preprocessing checks passed; YAML, shell examples and local documentation links checked.
• On sap4, 9,478 training and 499 evaluation records matched the previous input IDs and labels at S512 and S8192.
• LF main completed the two-step 8-GPU train/save smoke. Both adapter files matched the previous successful run's SHA-256 hashes.

The new artifact's inference retry ended with SIGKILL before serving requests. Earlier candidate-stack generation passed; this PR does not claim a successful new-artifact reload or clean-install release acceptance.