feat(skill): Add crypto-quant-ml with PurgedCV and Stationarity checks

[John-Swindell]

Context: Why this is needed Standard LLMs and generic coding agents struggle with the nuance of financial time-series, often defaulting to standard scikit-learn practices (like ShuffleSplit or MinMaxScaler) that introduce massive lookahead bias.

Furthermore, most financial ML context in training data is equity-centric (market hours, business days). This skill bridges the gap for Crypto-Native ML, enforcing strict academic rigor and adapting workflows for 24/7, high-volatility markets. It acts as a guardrail against common hallucinations that produce "great backtests" but failed production strategies.

Key Features & Methodology:

• Stationarity Enforcement: Automates Augmented Dickey-Fuller (ADF) tests to reject non-stationary predictions (predicting returns vs raw prices).
• Leakage Prevention (Purged K-Fold): Implements Lopez de Prado’s Purged/Embargoed K-Fold to respect temporal ordering and prevent training on testing outcomes.
• Crypto-Specific Logic: Enforces continuous calendars (no weekends/holidays), regime filtering (post-2020 institutional adoption), and dynamic universe selection to avoid survivorship bias.
• Anti-Pattern Detection: Explicitly forbids common pitfalls like Global Scaling (fit_transform on whole dataset) and Time-Based sampling for high-velocity assets.

Includes a full pytest suite in tests/ validating the stationarity logic and leakage detection.