Sunset the LEGACY and CURRENT dual model of includes

[eddelbuettel]

Following #475 last year, its eight sub-issues, and the related PRs and test runs, we accommodated the transition to a 'C++14 or later' demand from Armadillo while allowing many / most of all CRAN packages to continue to build.

As of today (18 March 2026) we still see some CRAN packages flagged for their C++ specification

> x <- tools::CRAN_check_details()
> x2 <- subset(x, Check == "C++ specification")
> table(x2$Package)

BayesGPfit        BSL      Eagle  EloChoice    fastmit   mixedCCA   multivar  RcppMeCab   robregcc  TFMPvalue 
         3          3          3          3          3          1          3         10          3          2 
> 

yet a full reverse-dependency run shows few to no failures. (For example, in the list above one package has its C++ specification only in src/Makevars.win which does not affect Linux-based tests.) It also appears that R now simply ignores the older C++ specification and just compiles under default compiler settings --- which is new enough for (Rcpp)Armadillo 15.0.0 or later and generally not throwing up any issues.

So with that we can look into relaxing this on returning to only including the current Armadillo sources.