speed up bounds-checked element access and cache vector lengths

[kevinushey]

Closes #1510.

The per-element bounds check added in #1310 was noticeably slowing down element access in tight loops, including sugar expressions. This PR keeps the check (and its warning) but makes it cheap:

• inst/include/Rcpp/vector/traits.h: the warning now lives in a noinline / cold helper (warn_index_out_of_bounds()), so tinyformat and the 8 KB message buffer are no longer inlined into every loop body that reads an element. proxy_cache (used by CharacterVector, List, ExpressionVector) now caches the vector length in update() rather than calling p->size() on every access.
• inst/include/Rcpp/vector/Vector.h: size() / length() return the length cached by update() instead of calling Rf_xlength() (an opaque call into libR) each time. update() already runs on every set__(), so the cache is refreshed whenever the underlying SEXP changes.
• inst/include/Rcpp/vector/Matrix.h, SubMatrix.h: cache the number of columns alongside the existing nrows, so ncol() / cols() (and MatrixRow::size()) no longer call Rf_isMatrix() + Rf_getAttrib() each time.

Benchmarks

R 4.6.1, arm64 macOS; the same functions compiled against master vs this branch. Results are identical between the two.

code	master	this PR	speedup	base R
sugar x + y (1e7)	43.9 ms	9.4 ms	4.7x	3.2 ms
sugar x * y + 2 * x - y (1e7)	61.4 ms	20.8 ms	2.9x	20.2 ms
for (i < x.size()) s += x[i] (1e7)	13.9 ms	5.6 ms	2.5x
CharacterVector loop, x[i] == NA_STRING (1e7)	42.4 ms	15.4 ms	2.8x
List loop, Rf_isNull(x[i]) (1e6)	7.4 ms	4.0 ms	1.9x
matrix loop with j < m.ncol() (1e6)	10.9 ms	1.1 ms	9.9x
MatrixRow loop with j < r.size() (1e6)	10.2 ms	0.7 ms	14.6x

Simple sugar expressions like x + y are still slower than with RCPP_NO_BOUNDS_CHECK, since the (never-taken) branch still blocks vectorization. Letting sugar read plain vector operands without a per-element check would close that gap, but that's a larger change to the sugar machinery, so I've left it for a follow-up.

Notes

• The cached lengths could only go stale if the underlying SEXP's length (or a Matrix's dim attribute) were changed in place behind Rcpp's back, e.g. via SETLENGTH() / R_resizeVector(). Nothing in Rcpp does this, and Matrix::nrow() already had the same caching behavior.
• The full tinytest suite passes locally with RunAllRcppTests=yes (1624 tests, 0 failures).

[Enchufa2]

Maybe the likely/unlikely macro technique used in the kernel helps too in these cases?

[Enchufa2]

I see there are C++ attributes too, but that means C++20.