Add Claude skills for corrrectness validation

[haok1402]

Quick regression check for code changes — run 15 training steps on both the base and feature branch, compare loss curves, and flag if anything diverged.

What's in here

• A Claude skill (.claude/skills/correctness-validation/) that walks the agent through the full workflow: figure out which models are affected, set up data, run both branches via git worktree, and compare the logs.
• Per-model setup and validation scripts for DeepSeek-V2-Lite (4 GPU) and Qwen3-30B-A3B (16 GPU), plus SLURM-aware launch scripts.
• compare.py parses stdout logs and checks cross-entropy loss, load-balance loss, and gradient norm step-by-step against a configurable tolerance (default 1e-3).
• save_interval and save_location in TrainingCfg are now optional (default None). Setting save_location without save_interval lets you load a pretrained checkpoint without ever writing one back — exactly what the validation scripts need to run a quick 15-step check without cluttering the checkpoint directory.

[gemini-code-assist]

Code Review

This pull request introduces a correctness validation framework to ensure code changes do not negatively impact training metrics. It includes a new Claude skill, model-specific setup and validation scripts for DeepSeek-V2-Lite and Qwen3-30B-A3B, and a utility to compare training logs. Additionally, the training configuration was modified to allow optional checkpointing. Feedback highlights a potential issue in the worktree setup documentation and recommends using explicit error handling instead of assertions for configuration checks.

[MasterJH5574]

LGTM, good to have the first skill 🎉