Add Ministral3 Eagle3 reproduction experiment

[andylolu2]

Why

This experiment checks whether the DeepSpec Eagle3 recipe can reproduce paper-level accepted-length results when adapted to mistralai/Ministral-3-3B-Instruct-2512. The final report is intentionally included even though the result is negative: the implementation works, but the 10k-subset run does not match the Qwen3-4B Eagle3 reference.

Summary

• Add Ministral3 Eagle3 model/trainer/evaluator support and 10k Open-PerfectBlend reproduction scripts.
• Add save-only checkpoints, incremental eval JSON writing, split eval task launching, and plotting utilities.
• Add a final markdown report with training curves and accepted-length comparison plots.

Reviewer guide

• Real logic changes: deepspec/modeling/eagle3/ministral3, deepspec/trainer/eagle3_trainer.py, deepspec/eval/base_evaluator.py, eval.py, and checkpoint handling in deepspec/trainer/base_trainer.py.
• Experiment/config changes: config/eagle3/eagle3_ministral3_3b.py and experiments/ministral3_3b_eagle3/*.sbatch.
• Report artifacts: experiments/ministral3_3b_eagle3/report.md and report_assets/ contain the final result; the report is the best entry point for reviewing the outcome.

Test plan

• Ran Python syntax checks for eval.py and shell syntax checks for the split eval sbatch.
• Ran smoke data/cache/train jobs for Ministral3 support before the full run.
• Completed the 3000-step DeepSpec Eagle3 training run and final 9-task eval. Final eval JSON: /mnt/vast/runs/andy/deepspec_ministral3_3b_eagle3/eval_10k_step3000_metrics.json.
• Regenerated report plots from TensorBoard and the completed eval JSON.