address review findings on the windows onetbb build

[kevinushey]

Follow-up to #269, which was merged before these review findings were applied.
Two are behaviour fixes; the rest tighten things #269 introduced.

Fixes

configure.R: derive the TBB library names whenever an Rtools tree is adopted.
#269 gated the whole Windows block on oneapi/ being present, but that block
does two jobs: deciding whether to adopt the tree, and deriving TBB_NAME /
TBB_MALLOC_NAME from the archive names. Gating both broke an explicitly
configured legacy TBB_LIB: the names fell back to tbb / tbbmalloc, so
PKG_LIBS emitted -ltbb against a directory holding libtbb_static.a. The
_static alternation in the pattern exists precisely for this, and both the
legacy pkgLibs branch and buildTbbStub's "re-export the static tbb library
wholesale" branch still contemplate a non-oneTBB Windows TBB. An explicitly
configured TBB is now honoured whatever its vintage; auto-discovery next to
gcc still requires oneTBB.

configure.R: don't let the cmake probe throw. The probe now runs on
Windows, where the whole point is degrading to tinythread -- but
system(..., intern = TRUE)[[1L]] raises a subscript error if CMAKE points
at nothing runnable, and numeric_version() raises on output it can't parse.
Either turned the intended fallback into a hard configure failure. Any probe
failure is now treated as "no usable cmake": non-fatal on Windows, still a
hard error off it.

R/tbb.R + downstream check: the -ltbb fallback isn't free. The ordering
rationale is right, but the comment presented the fallback as pure upside. The
stub links -ltbb12 itself, so it embeds its own copy of the oneTBB runtime; a
package resolving some symbols from RcppParallel.dll and the rest from the
stub would straddle two independent schedulers, and an observer registered with
one would never fire for arenas owned by the other. In practice most of the
archive is pulled into RcppParallel.dll transitively (arena.cpp references
observer_list), but the failure mode changed from a link error to a silent
correctness bug. The comment now says so, and tbb-downstream-check.R asserts
via objdump -p that nothing lands on the stub -- which also answers one of
#269's own open CI questions.

GNU.cmake: probe with an empty file, not the null device. gcc -c NUL
was plausible but unverified and depends on gcc accepting a device name as an
input file. A file(WRITE ...) probe works identically everywhere and drops
the WIN32 branch entirely.

install.libs.R: don't log the Windows static build as a failure.
.install.libs's shlibPattern still matches ^tbb.*\.dll$ on Windows, so a
bundled Windows build logged no tbb runtime libraries found in 'tbb/build/lib_release'. Benign -- the archive is linked into
RcppParallel.dll -- but it reads like a failure in exactly the logs someone
consults when this goes wrong.

Smaller items

• windows-11-arm added to the downstream matrix: RcppParallelLibs()
  changed for every Windows platform, and arm64 has both its own vendored
  patch and itt_notify compiled out.
• Corrected the required cmake version in the error message (>= 3.5, matching
  the check and SystemRequirements).
• Renamed useBundledTbb to buildBundledTbb in configure.R, so it no
  longer collides with the useBundledTbb() function in install.libs.R.
• Restored the trailing newline in observer_proxy.cpp.
• rstan.yaml pins against DESCRIPTION rather than a .9000 suffix, so it
  survives development version bumps.
• actions/checkout@v3 -> v4, and consistent ${R_ARCH_BIN} in the
  downstream Makevars.
• Regenerated patches/legacy_tbb_abi.diff and patches/mingw_gnu_cmake.diff
  to match the amended sources.

Verification

• Configure decision matrix -- simulated 10 toolchains by sourcing the real
  configure.R into a mocked environment. Rtools43+ output is byte-identical
  to build the bundled onetbb on windows when rtools has no onetbb #269, so its compatibility claim still holds. Confirmed against the
  pre-fix code that both bugs were real: the explicit legacy TBB_LIB case
  produced TBB_NAME: tbb, and an unparseable cmake version aborted configure
  with invalid version specification.
• macOS -- full R CMD INSTALL plus all 9 test files; the downstream check
  passes end-to-end (CXX_STD confirmed honoured).
• CMake probe -- tested standalone: the empty probe file compiles without
  "no input files", and the unparseable-version fallback prevents the
  math(EXPR) failure.
• Import-table parsing -- unit-tested against synthetic objdump output,
  including case-insensitivity and the last-block boundary.
• Patches -- all five reverse-apply cleanly against the vendored tree.

What still needs CI

Unchanged from #269: none of the Windows behaviour is testable locally. Two
additions here can legitimately turn CI red and are the ones to watch --
the windows-11-arm downstream job, and the import assertion, which fails if
anything genuinely does resolve from the stub. Both surface information rather
than assume it.

Note patches/windows_arm64.diff does not reverse-apply (it references
src/tbb/build/, which is .Rbuildignored). That predates #269 and is left
alone here.