Expand README with end-to-end data engineering, model, and evaluation docs

[Sam-v6]

Motivation

• Clarify the end-to-end workflow so users understand how raw tracking data is transformed into model-ready tensors and evaluation artifacts.
• Document the responsibilities of the core scripts to make reproduction and extension easier (src/create_features.py, src/train_transformer.py, src/generate_predictions.py, src/process_predictions.py).
• Explain how model outputs (probabilities) can be used for additional evaluation like ROC/AUC beyond the existing confusion matrix and accuracy plots.

Description

• Added a new README section that walks through the data engineering and feature creation pipeline implemented in src/create_features.py and src/clean_data.py, including loading, cleaning, subsampling, augmentation, tensorization, and saved tensor artifacts.
• Documented the transformer architecture in src/models/transformer.py and the training flow in src/train_transformer.py, including data loading, optimizer/loss choices, mixed-precision inference, saving the best model, and optional Ray Tune HPO behavior.
• Described the streaming inference process in src/generate_predictions.py, how per-frame tensors are constructed, and where per-frame zone_prob/man_prob and prediction CSV/Parquet outputs are written.
• Documented post-processing in src/process_predictions.py, including accuracy-vs-time plots, per-play animations, confusion matrix generation, and notes on deriving ROC/AUC from saved probabilities.

Testing

• This is a documentation-only change; no code behavior was modified and no automated tests were run for this PR.
• The change was committed locally and staged as a docs update (README.md) with no runtime validation steps executed.

---

Codex Task