feat: add exact Shapley risk attribution

[RichaShiny]

Adds an exact Shapley-value attribution layer for RiskForge's nonlinear cascade model.

What this adds:

• exhaustive coalition evaluation across ETH, BTC, and SOL shocks
• exact Shapley decomposition with no sampling approximation required for three assets
• attribution for first-order exposure, cascade-aware exposure, cascade amplification, bad debt, secondary liquidations, and endogenous price decline
• explicit efficiency residuals showing that allocated Shapley contributions reconcile to the full portfolio change relative to the no-shock baseline
• paired stochastic Shapley attribution across identical simulated market draws
• tail-conditioned Shapley summaries for cascade exposure and bad debt
• deterministic regression coverage for efficiency, zero-shock invariance, isolated-asset behavior, replay reproducibility, tail aggregation, and invalid inputs

This complements the existing leave-one-asset-out attribution view. Leave-one-out effects remain useful as direct counterfactual sensitivity measures, while Shapley values provide an additive allocation of nonlinear interaction effects across collateral assets.

Please review and merge manually.