Add assume_PSD=True to ex_ante_tracking_error's quad_form

[shlokkvaishnav]

What this fixes

objective_functions.py has four cp.quad_form(...) call sites over a covariance matrix. Three of them (portfolio_variance, sharpe_ratio, quadratic_utility) pass assume_PSD=True, which skips cvxpy's own numerical PSD-certification check on the matrix — the right call, since cov_matrix is PSD by construction even when it's numerically ill-conditioned (small sample size relative to number of assets, near-duplicate assets, etc.).

ex_ante_tracking_error was the one outlier missing it, despite taking the same kind of cov_matrix argument used the same way. This is the exact failure class reported in #631 (cvxpy note: ... trying to certify that a matrix is positive semi-definite ... replace the matrix A by cvxpy.psd_wrap(A)), for a user's own quad_form-based objective modeled on this codebase's pattern.

Honesty about verification

I was not able to force this exact PSD-certification error against the current cvxpy (1.9.2) with any covariance matrix I constructed (including exactly-singular, ill-conditioned, and rank-deficient ones) — cvxpy appears to have become more lenient about this over time. So this isn't a "fails on main, passes with fix" reproduction; it's a defensive consistency fix that matches the established pattern used by the sibling objective functions in this same file, and directly addresses the failure mode described in #631.

Test log

$ pytest tests/test_objective_functions.py -v
...
12 passed in 6.30s

$ pytest tests/test_efficient_frontier.py -q
75 passed, 6 warnings in 39.46s

$ black --check pypfopt/objective_functions.py tests/test_objective_functions.py
All done! ✨ 🍰 ✨
2 files would be left unchanged.

Added test_ex_ante_tracking_error_assumes_psd confirming the numeric output is unchanged for a well-conditioned matrix (i.e. the fix is behavior-preserving in the normal case), alongside the existing test_ex_ante_tracking_error / test_ex_ante_tracking_error_dummy tests which both still pass unmodified.