Support right-censored / top-coded counts (e.g. 1, 2, 3+) in zero-truncated models

[BERENZ]

Extend the zero-truncated families to handle right-censored (top-coded) counts, where the count y is observed only in grouped form such as 1, 2, 3+ (the top category lumps all y ≥ c) rather than the full 1, 2, 3, 4, … .

This is the concrete, count-specific version of the "right-censored models" line in #10.

Motivation

Capture–recapture / unseen-species data are frequently reported in top-coded form — e.g. "identified once, twice, or three-or-more times". The Chao-type estimators (chao, oichao) use only the low counts (f₁,f₂ or f₂,f₃) and are therefore already robust to top-coding, but the full-likelihood parametric families (ztpoisson, ztgeom, ztnegbin, the OI/ZT and Hurdle variants, …) currently assume every count is observed exactly and cannot be fit to top-coded data.

What's needed

• Allow the response to encode a censoring threshold c, so the data are (f₁, …, f_{c−1}, f_{c+}) with f_{c+} = #{units with y ≥ c}.
• In the (zero-truncated) likelihood, use the exact pmf p₊(y) for the fully-observed cells y < c and the tail / survival probability P(Y ≥ c | Y ≥ 1) = 1 − Σ_{y=1}^{c−1} p₊(y) for the censored top cell. (The families already carry the pmf and a CDF — used internally in their simulate methods — so the survival term is readily available.)
• Propagate to deviance / residuals and confirm the Horvitz–Thompson / popSizeEst() machinery works once parameters are estimated from censored data.

Scope / notes

• Primarily benefits the parametric ZT families; chao/oichao are unaffected (they ignore high counts by construction).
• Related but distinct from the binomial-with-known-T base (Add Binomial-with-known-T base for OIZT/ZTOI (Boehning & Friedl 2024, Sec. 5) #64), which is bounded support (y ≤ T); this is censoring of an otherwise-unbounded count.
• Likely a controlModel / response-encoding option plus a family survival/tail hook, rather than new families.

Acceptance

• Fit a ZT family to top-coded data (e.g. {1, 2, 3+}) and recover sensible parameter + population-size estimates; unit tests on a synthetic top-coded sample; documentation.