Carry the model's weights and tie handling into ggforest_subgroup() refits

[kassambara]

ggforest_subgroup() gets each subgroup hazard ratio by refitting the Cox model on
that subset, but the refit was a plain coxph(): the case weights, tie handling and
clustering of the supplied model were dropped. A weighted model reported unweighted
subgroup estimates, and a clustered model drew intervals from the naive variance.

Those settings are now carried into every refit. Weights and clustering ride along as
data columns so subsetting the data subsets them too, and both are read from the fitted
object rather than re-evaluated from its call.

The interaction test is affected the same way, and it cannot simply be refit weighted:
coxph() attaches a robust variance to any non-integer weighted fit, and anova()
declines to compare likelihoods for such a fit because their differences are no longer
chi-square. A fit carrying a robust variance now gets a Wald chi-square on the
treatment-by-subgroup coefficients instead; everything else, integer weights included,
keeps the likelihood-ratio test.

A data whose rows cannot be matched to the fit is now refused when weights or
clustering have to be carried across.

library(survminer)
library(survival)

cc <- colon[colon$etype == 2 & colon$rx %in% c("Obs", "Lev+5FU"), ]
cc$rx    <- droplevels(cc$rx)
cc$sex   <- factor(cc$sex, labels = c("Female", "Male"))
cc$node4 <- factor(cc$node4, labels = c("<= 4 nodes", "> 4 nodes"))

set.seed(9)
cc$w <- runif(nrow(cc), 0.2, 3)          # inverse-probability weights

m <- coxph(Surv(time, status) ~ rx + age, data = cc, weights = w)

ggforest_subgroup(m, data = cc, treatment = "rx",
                  subgroups = c(Sex = "sex", Nodes = "node4"))

Before — the refits ignore the weights:

After — the overall row reproduces exp(coef(m))["rxLev+5FU"] = 0.70, and the female
subgroup moves from an apparent 13% reduction to exactly no effect:

Models specifying none of weights, ties, robust or clustering are unchanged.