Add inequality constraints to Gaussian Process trainer

[dschwen]

Reason

Enforce inequality constraints when training Gaussian processes.

Design

Three mechanisms for enforcing inequality constraints on GP predictions:

• Link functions (log, logit): exact constraints via output warping; GP trains in unconstrained latent space and predictions are inverted back to physical space via the inverse link.
• Derivative observations: monotonicity constraints via virtual observations that augment the covariance matrix with cross-covariance blocks between function values and directional derivatives.
• Penalty constraints: soft bounds added to NLML loss/gradient during Adam hyperparameter tuning.

Impact

Added capability