Add online EWA and shrinkage covariance/precision estimators

[MaxHalford]

What

Adds a family of online covariance estimators to river.covariance, reimplemented from the precise package following River conventions (dict-native, Narwhals mini-batches), per discussion #1884. Only genuinely-online methods are included — nothing that lazily inverts a stored covariance on read.

New estimators (river/covariance/ewa.py)

• EwaCovariance — exponentially weighted (RiskMetrics-style) covariance. Diagonal matches stats.EWVar, off-diagonals match stats.EWCov. For non-stationary streams whose relationships drift over time.
• LedoitWolfCovariance / OASCovariance — data-driven shrinkage of the EWA covariance towards a scaled identity (Ledoit-Wolf 2004 / Chen et al. 2010 intensities). For high-dimensional / few-sample regimes where the raw covariance is noisy or singular.
• ShrunkCovariance — fixed-intensity shrinkage with a finance-friendly constant-correlation target (or identity).
• EwaPrecision — exponentially weighted precision (inverse covariance) maintained online via a forgetting-factor Sherman-Morrison update. Genuinely online (O(d²)/step, never inverts explicitly); the recency-weighted counterpart of EmpiricalPrecision.

Supporting additions

• stats.EWCov — exponentially weighted covariance primitive (bivariate counterpart of stats.EWVar), composed from EWMeans so the convention matches exactly.
• datasets.SP500Stocks — daily returns (in %) of ten large-cap S&P 500 stocks across diverse sectors (2013–2018, 1,257 trading days), used in the docstring examples. Bundled sp500.csv.gz.
• Guards SymmetricMatrix.__repr__ against empty (unfitted) matrices (previously raised ValueError).

Design notes

• Internals are array-backed with a feature→index map (the same pattern as the existing EmpiricalPrecision) behind a dict-native public interface (update, update_many, matrix, __getitem__).
• The EWA convention reuses stats.EWMean/EWVar so the diagonal and off-diagonals are exactly the existing scalar EW statistics.
• No precision equivalents for the shrinkage estimators: shrinking toward a scaled identity is a full-rank perturbation that can't be tracked by rank-one inverse updates, so a "shrunk precision" would require inverting on read — exactly the lazy approach this work excludes.

Tests

• river/covariance/test_ewa.py: EWA vs independent numpy reference + stats.EWVar/EWCov; update_many ≡ single-update loop across all estimators and across pandas/polars backends; shrinkage vs sklearn.covariance oracle; EwaPrecision vs numpy oracle and P @ S ≈ I; PSD/symmetry invariants; pickling; empty-repr.
• All doctests use the real SP500Stocks data. uv run mypy and ruff are clean.

🤖 Generated with Claude Code

[MaxHalford]

@microprediction this PR pulls some of precise's methods into River. I'm very grateful for this gift!