cusna 0.1.0

First release: a self-contained native engine (C ABI over CUDA kernels and C++ host logic) for the simulation and estimation of network models, callable from R without a Python runtime. The CRAN build is CPU-only from source; the GPU path compiles when a CUDA toolkit is detected at configure time.

• SAOM (the RSiena model family). saom_data(); effect constructors (cusna_effect(), cusna_beh_effect(), cusna_rate_effect(), cusna_interaction()); the full Method-of-Moments estimator mom_estimate() / mom_control() returning a cusna_fit object with summary(), coef(), vcov(), and as.data.frame() methods; behavior co-evolution, composition change, conditional and unconditional estimation; multi-network co-evolution (saom_multinet_data(), mom_estimate_multinet()); and a siena07() simulation backend (cusna_fran()). Data preparation, effect preprocessing, and moment targets are validated bit-for-bit against the reference implementation; estimates agree within simulation standard errors.
• ERGM. Sufficient statistics (ergm_stats()), a TNT sampler (ergm_simulate()), pseudo-likelihood (ergm_mple()), and MCMC maximum likelihood (ergm_mcmle()), matching ergm::ergm() on benchmark models.
• Temporal ERGM. tergm_mple() (pooled MPLE with block bootstrap, matching btergm), temporal simulation (tergm_simulate()), and the separable stergm_cmle() (formation/persistence, matching tergm CMLE).
• ALAAM. alaam_mple() (exactly reproducing the corresponding glm), alaam_mcmle(), and a Gibbs simulator (alaam_simulate()).
• Low-level host statistics (cusna_network_stats(), cusna_behavior_stats(), cusna_gof_distribution()) reproduce RSiena targets on public datasets to machine precision.