Notebook: add checkpoint resume, CPU threading, and finer checkpoints to AdultIncome long-run tracking

[charlesmartin14]

Motivation

• Improve robustness for long-running training by enabling resume-from-checkpoint behavior so interrupted runs can continue without restarting from zero.
• Improve CPU utilization and platform awareness by auto-detecting available threads and Apple Silicon to tune XGBoost threading.
• Increase checkpoint frequency and finer-grained training chunks to reduce potential wasted work between checkpoints.

Description

• Added import os and import platform and compute CPU_THREADS and IS_APPLE_SILICON, and set nthread in params with explanatory comment for macOS ARM.
• Reduced CHUNK_SIZE from 50 to 25, changed CHECKPOINT_EVERY_STEPS from 2 to 1, and introduced RESUME_FROM_CHECKPOINT = True.
• Implemented resume logic that reads results_path_csv and latest_model_path, validates that the last saved round is divisible by CHUNK_SIZE, loads the xgb.Booster to continue training, and sets start_step accordingly.
• Ensure checkpointing persists metrics to CSV and Feather (with try/except), saves per-round model and latest_model_path, and added diagnostic prints and a final results_df fallback to ensure results are available.

Testing

• Executed the notebook initialization cells in Colab to verify configuration prints for TOTAL_ROUNDS, CHUNK_SIZE, CHECKPOINT_EVERY_STEPS, RESUME_FROM_CHECKPOINT, and CPU_THREADS, and the prints completed successfully.
• Ran the training loop cells for several steps to verify checkpoint saving to results_path_csv and model files and to confirm results_df.tail() returns without error.
• Verified resume path by simulating existing results_path_csv and latest_model_path, and confirmed the code loads the booster and resumes from the expected round without exceptions.
• All automated notebook cell executions used during testing completed successfully with no uncaught exceptions.

---

Codex Task