Add differentiable CVaR to UQPCEGroup and MultiUQPCEGroup

[bdphilli]

Adds conditional value at risk (CVaR) as a differentiable statistic alongside the existing confidence intervals, for use as an optimization objective or constraint.

API

A new compute_cvar option (default False) on UQPCEGroup/MultiUQPCEGroup adds cvar_lower/cvar_upper outputs next to ci_lower/ci_upper, with the same tail convention: significance=0.05 gives the mean of the worst 2.5% tail, so each CVaR pairs with its CI. The existing use_tanh_ci flag selects the implementation, exactly as it does for the CI.

Exact form (use_tanh_ci=False)

CVaRComp, a JAX component computing the Rockafellar-Uryasev form

CVaR_upper = VaR + mean((x - VaR)+) / (alpha/2)
CVaR_lower = VaR - mean((VaR - x)+) / (alpha/2)

on the resampled responses, with the same aleatory/epistemic reshape and max/min aggregation as CDFComp. Well defined for repeated (discrete) sample values; reduces to the conditional tail mean for continuous samples.

Smoothed form (use_tanh_ci=True)

CVaRGroup solves the tanh-smoothed confidence interval with the same CDFResidComp/BalanceComp machinery as CDFGroup, then CVaRTailComp evaluates the Rockafellar-Uryasev form with a softplus hinge. Since softplus >= relu, the smoothing bias is always outward -- the upper CVaR is overestimated and the lower underestimated -- mirroring the conservative bias of the tanh CI. Bias scales as omega^2 * density / tail_frac. KS aggregation across epistemic curves as in CDFGroup.

Verification

• Exact form matches numpy references to 1e-6 and the analytic standard-normal CVaR.
• Analytic totals match finite difference to ~1e-8 in both modes, including through the Newton balance.
• The conservative bias direction is asserted in the tests.
• Ten new unit tests under test_suite/test_uqpce/mdao/cvar/, mirroring the cdf test layout; the full mdao suite passes.

Example

examples/GMM/objectives_comparison_example.py optimizes the same two-variable analytical problem (built on the GMM example's uncertainty model) under each statistic -- deterministic, mean, mean_plus_var, ci_upper, cvar_upper -- and each design provably minimizes its own column of the printed statistics table. The model is quadratic in the uncertain inputs so the order-2 PCE is exact, and the CI-vs-CVaR distinction is interpretable: CVaR shifts exposure away from the unbounded multimodal input and onto the bounded one. Writes a design-space/CDF/PDF comparison figure; reproducible run-to-run (relies on the seeding fix from #25).