Add BlockCovariance: memory-efficient block-diagonal online covariance (O(p·b) state)

[microprediction]

Summary

Adds BlockCovariance — a memory-efficient block-diagonal online covariance addressing the streaming-memory angle of #47.

SchurCovariance(gamma=0) produces a block-diagonal estimate but keeps the full p×p EWA covariance in state and zeroes cross-block entries only at read time. BlockCovariance instead tracks only the within-block (contiguous n_blocks) recency-weighted covariances:

• State and per-step cost ~ O(p·b) (block size b), not O(p²) — it never allocates the dense matrix while streaming; it densifies only on a covariance_ request.
• PD by construction — a block-diagonal of positive-definite blocks is positive-definite, so no global O(p³) eigen-projection; the precision is block-wise too.
• It is the memory-light gamma=0 (block-diagonal / composite) member of the Schur family — useful at high p where the dense covariance is expensive or impossible to hold (the regime Block-streaming Schur pseudo-likelihood: avoid materializing the full p×p covariance #47 targets), and a natural generalization of DiagonalCovariance (the b=1 case).

Tests

• Covered by the parametrized all_estimators() conformance suite (contract, fit == stream, state roundtrip, real-JSON state).
• New tests/test_block_covariance.py: block-diagonal structure (off-block entries exactly zero), positive-definiteness, sub-quadratic state (sum of block elements < p²), JSON-serializable state that roundtrips, and blocks matching a within-block EwaCovariance.
• ruff, mypy, and full pytest pass locally.

Notes

This is the block-diagonal case of #47. A general bounded-conditioning (Vecchia-style) streaming variant that keeps limited cross-block coupling without the dense matrix is a natural follow-up. Complementary to #48 (SchurLedoitWolfCovariance, the analytic cross-block damping).