Add cross-sectional partner-dependence predictive checks

[Pascal-Kueng]

Adds descriptive checks of partner variation and dependence in cross-sectional dyads. The workflow predicts responses with conditional-model random effects set to zero, simulates complete datasets with new random-effect blocks, subtracts the same predictions from observed and simulated responses, and compares paired summaries in tables and plots.

Scope and implementation

• Supports unweighted glmmTMB models without zero inflation for Gaussian, Poisson, negative binomial 1 and 2, Tweedie, Gamma, and beta responses. Prediction and simulation use the fitted link; the wrapper checks the family and requires finite scalar responses.
• Provides model-centred or raw checks, with role-specific or exchangeable summaries. Undefined simulated statistics are counted explicitly.
• Simulation results contain observed_response, simulated_responses, predicted_response, and model_frame. Model and simulation metadata live in attr(x, "dyadMLM"); simulation counts come from matrix dimensions.
• Preserves fitted-row alignment and complete random-effect blocks, restores simulation settings, and preserves the caller's random-number state when a seed is supplied. Print methods retain their internal documentation keyword.
• Keeps fitted parameters and study design fixed. These are descriptive predictive references, without refitting or parameter uncertainty; reference intervals and observed quantiles are not confidence intervals or calibrated p-values.

ILD checks, adapters for other response formats, and vignette integration remain follow-up work. Developer notes and the vignette draft are shortened and aligned with the public functions.

Validation

• Built-source R CMD check --no-manual --ignore-vignettes: 0 errors, 0 warnings, 0 notes.
• 1,566 passing test assertions, with no failures, warnings, or skips.
• Actual Gaussian/log, Poisson/sqrt, and beta/probit fits match glmmTMB predictions and simulations, alongside the existing family, alignment, state-restoration, and plotting tests.
• Generated help checked; developer simulation study rendered as an execution smoke check with one outer repetition and 20 inner simulations.