autorelevate 0.1.0

• Initial CRAN submission.
• Implemented the autorelevated family of distributions (PDF, CDF, survival, hazard, exact quantile function via Lambert W_-1, and random generation) for ten two-parameter baseline lifetime distributions: Weibull, Lomax, Burr XII, Gompertz, Log-Logistic, Chen, Exponentiated Exponential, Power Lindley, Log-normal, and Gamma (dautorelevate(), pautorelevate(), sautorelevate(), haautorelevate(), qautorelevate(), rautorelevate()).
• Added Maximum Likelihood, Maximum Product of Spacings, Least Squares, Weighted Least Squares, and Cramer-von Mises estimation methods via fit_autorelevate(), reporting AIC, BIC, CAIC, and HQIC for MLE method.
• Added fit_all_methods() and compare_families() for comparing estimation methods and all ten members of the autorelevated family.
• Added ttt_plot() for Total Time on Test plots (Aarset, 1987), a standard tool to determine the hazard function shape.
• Added print(), summary(), and plot() methods for autorelevate_fit objects, and a print() method for autorelevate_compare objects.
• Added a bundled real dataset, bladder_cancer (Lee & Wang, 2003; n = 128 bladder cancer remission times), with no dependency on external data packages. The MLE fit of the Autorelevated Weibull family to this dataset reproduces the published results of Dileep Kumar, Shabeer, and Sankaran (2025) to within optimizer tolerance.
• Added an internal input-validation layer used by every exported function, giving clear error messages for invalid dist, p1, p2, or data instead of silent NaN/Inf propagation, and a warning (rather than silent truncation) when non-finite or non-positive observations are dropped from data.
• Made optimization in fit_autorelevate() robust via a data-adaptive starting-value grid search and an automatic Nelder-Mead/restart fallback if the primary optimizer fails to converge.
• Fully vectorized the internal Lambert W_-1 solver (.lambert_w_minus1()) for performance at scale, with a convergence check and warning.
• Fixed a rare edge case in the internal Lambert W_-1 solver where an extreme quantile request (p within 1e-10 of 1) could silently return a slightly inaccurate result; verified against independent high-precision root-finding to now be accurate to machine precision across the full domain.