Optimize cluster-robust vcov aggregation

[yangzichao]

Benchmark (200k x 10, 2k clusters): baseline 0.0071s mean, new 0.0068s mean (~4-6% faster).

Methodology references (required if estimator / math changes)

• Method name(s): Cluster-robust sandwich variance estimator (CRSE), HC1 small-sample adjustment
• Paper / source link(s): N/A (implementation detail; estimator unchanged)
• Any intentional deviations from the source (and why): None; only aggregation strategy changed (pandas groupby → NumPy add.at) in the NumPy fallback.

Validation

• Tests added/updated: None
• Backtest / simulation / notebook evidence (if applicable): Micro-benchmark run locally (200k rows, 10 regressors, 2k clusters).

Security / privacy

• Confirm no secrets/PII in this PR: