Extends the hierarchical portfolio module with two additional methods:
HERCOpt - Hierarchical Equal Risk Contribution (Raffinot 2018)
 Equal Risk Contribution (ERC) weights at every level of the hierarchy
 instead of the inverse-variance weights used by HRP. ERC accounts for
 within-cluster correlations, producing more balanced risk allocation
 when intra-cluster correlations are high. Implemented via cyclical
 coordinate descent (Roncalli 2013).
NCOpt - Nested Cluster Optimization (Lopez de Prado 2019)
 Two-level nested procedure: within-cluster optimization followed by
 meta-portfolio optimization across clusters. Supports three objectives
 at each level - min_variance, erc, and equal-weight - giving 9
 combinations. The meta-covariance is computed analytically from the
 full covariance matrix, so no return history is required.
Both classes:
 - Share the same API as HRPOpt (fit from returns or cov_matrix)
 - Export portfolio_performance() with expected return and Sharpe ratio
 - Are exported from pypfopt.__init__ and __all__
 - Accept all scipy linkage methods
Also adds two private helpers:
 - _erc_weights_ccd: ERC via multiplicative CCD (Roncalli 2013)
 - _min_var_weights: long-only min-variance with analytic + SLSQP fallback
53 tests across 9 test classes; all existing HRP tests still pass.
References
----------
Raffinot, T. (2018). Hierarchical clustering-based asset allocation.
Journal of Portfolio Management, 44(2), 89-99.
Lopez de Prado, M. (2019). A Robust Estimator of the Efficient Frontier.
SSRN Working Paper. https://ssrn.com/abstract=3469961
Roncalli, T. (2013). Introduction to Risk Parity and Budgeting.
CRC Press.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

... CCD
The previous implementation used Roncalli's multiplicative update, which
diverges when the covariance matrix has negative off-diagonal entries because
some risk contributions become negative during iterations and are incorrectly
clipped before taking the square root.
Replace with the Spinu (2013) cyclical coordinate descent, which minimises
the unconstrained convex objective
 f(w) = (1/2) w'Σw − (1/n) Σi log(wi)
by solving the exact one-dimensional sub-problem at each coordinate:
 Σii·wi2 + (Σw − Σii·wi)·wi − (1/n) = 0 → positive root
Crucially, the weights must NOT be normalised between coordinate updates;
normalisation happens once, at the end of each full pass over all assets.
This change fixes test_equal_risk_contributions (which uses a covariance
matrix with negative off-diagonal entries) and leaves all HERC / NCO tests
unchanged (53/53 pass), since the algorithm is equivalent for PD matrices
with the results scaled consistently.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>