per-element bounds check and uncached lengths slow down vector access and sugar #1510

The bounds check added to vector element access in #1310 has a noticeable cost in tight loops, including sugar expressions (which read every operand through Vector::operator[]).

Two things contribute:

1. r_vector_cache::check_index() and proxy_cache::check_index() (inst/include/Rcpp/vector/traits.h) call Rcpp::warning() inline. That pulls tinyformat plus an 8 KB stack buffer into every loop body that reads an element, which appears to block inlining and vectorization even though the branch is never taken.
2. proxy_cache::check_index() calls p->size() on every access, and Vector::size() / length() call Rf_xlength() on every call. With current R headers this is an opaque function call into libR, so for (R_xlen_t i = 0; i < x.size(); i++) makes a libR call per iteration -- even though r_vector_cache already stores the length. Similarly, Matrix::ncol() calls Rf_isMatrix() + Rf_getAttrib() on each call (and MatrixRow::size() goes through it), while nrow() is already cached.

Micro-benchmarks (R 4.6.1, arm64 macOS, 1e7 elements, current master):

Proposed fix (no behavior change; the out-of-bounds warning is kept):

• Move the warning into a separate noinline / cold helper and mark the check as unlikely.
• Have Vector::size() return the length cached by update() (which already runs on every set__()), and cache the length in proxy_cache too.
• Cache the number of columns in Matrix alongside nrows.

A possible follow-up would be letting sugar read plain-vector operands without a per-element check (as the old Fast extractor did), but that is a larger change to the sugar machinery.