Add end-to-end offline tutorial notebook

[jsp2195]

Motivation

• Provide a single, runnable Jupyter tutorial that demonstrates the repository's full offline workflow end-to-end (tokenizer → pretrain → eval → SFT → optional DPO → inference) using only local toy data and in-repo APIs.
• Make it easy for users to exercise the stack on CPU or GPU with modest default runtimes and clear knobs to shorten/lengthen runs.

Description

• Add exactly one file: notebooks/end_to_end_tutorial.ipynb, containing labeled sections for Setup, Tokenizer training, Pretraining (tiny GPT), Evaluation (perplexity + probes), SFT, optional DPO (toggle), and a minimal greedy generation demo.
• The notebook performs robust repo-root detection, imports from src/llmstack modules, writes artifacts under artifacts/ and runs/, sets a deterministic seed via seed_everything, and auto-selects device/precision (bf16/fp16 on CUDA, fp32 on CPU).
• Runtime knobs (PRETRAIN_STEPS, SFT_STEPS, DPO_STEPS, RUN_DPO) are centralized and modest by default (pretrain ~80 steps, SFT ~30, DPO ~20) to keep CPU runs feasible, and all paths are relative to the repository root.
• The notebook is offline-first and does not download external models or datasets, and it uses the repo's tokenizer wrapper, DataModule, Trainer, evaluate_perplexity, run_probes, and post-training helpers where appropriate.

Testing

• Confirmed the notebook file was written and is valid JSON by parsing it with a short Python snippet that printed nbformat, nbformat_minor, and the cell count (script executed successfully).
• Displayed the notebook contents to verify the expected sections and that the primary runtime knobs and output paths (artifacts/ and runs/) are present.
• No automated unit tests in tests/ were modified, and CI configuration was left untouched by this change; the notebook is intended for local interactive execution (offline) and was validated for offline-only behavior in the added cells.

---

Codex Task