build: enable -Werror on TA translation units (opt-in via TA_WERROR)

[evaleev]

Summary

Add an opt-in TA_WERROR CMake option that promotes compiler warnings
to errors on TiledArray's own translation units (library, tests,
examples), default OFF, not propagated to consumers of the installed
tiledarray target. Wire it on in both CIs.

Modeled after the MADNESS counterpart, m-a-d-n-e-s-s/madness#693.

What changes

CMake — opt-in TA_WERROR, default OFF.

• New cmake/modules/TiledArrayWarnings.cmake owns an INTERFACE
  library tiledarray_internal_warnings. With TA_WERROR=ON it
  carries -Werror (CXX/C/HIP) and -Xcompiler=-Werror (CUDA), gated
  on GNU/Clang/AppleClang/IntelLLVM (warns otherwise).
• src/CMakeLists.txt: pulls compile options off the warnings target
  via $<TARGET_PROPERTY:tiledarray_internal_warnings,INTERFACE_COMPILE_OPTIONS>
  rather than a real link, so the (intentionally non-installed)
  warnings target does not get pulled into the tiledarray export
  set.
• add_ta_executable: PRIVATE-links the warnings target to every
  in-tree executable, covering ta_test and all examples/.
• examples/scalapack/CMakeLists.txt: same wiring for the one raw
  add_executable.

Verified by inspecting the generated tiledarray-targets.cmake: zero
Werror occurrences, tiledarray_internal_warnings absent from every
INTERFACE_LINK_LIBRARIES.

TA_WERROR=ON implies MADNESS_WERROR=ON via
FindOrFetchMADWorld.cmake, so MADworld TUs built as part of TA's
FetchContent tree are covered. This requires a MADNESS pin that
includes the MADNESS_WERROR option, so TA_TRACKED_MADNESS_TAG is
bumped to the head of m-a-d-n-e-s-s/madness#693 (7d8aaf9d5); this
PR is blocked on that one landing.

Scope of -Werror intentionally excludes:

• Bundled FetchContent code (parsec, BTAS, Umpire, range-v3, blaspp,
  lapackpp) — those targets are not built via add_ta_*.
• python-tiledarray — pybind11-generated code, built via
  pybind11_add_module.

range-v3 marked as SYSTEM include. range-v3 self-triggers
-Wdeprecated-declarations from compressed_pair.hpp using its own
deprecated ranges::compressed_tuple alias. Header-only and no
include-order risk against TA's headers, so it's carved out from the
top-level CMAKE_NO_SYSTEM_FROM_IMPORTED=TRUE.

Source-level warning fixes:

• Drop unused locals / type aliases / lambda captures in
  array_impl.h, binary_engine.h, cont_engine.h
  (-Wunused-{variable,local-typedef,lambda-capture}).
• Mark header-defined free functions / type aliases in einsum/string.h
  and tensor/kernels.h [[maybe_unused]].
• Replace umbrella range/v3/{algorithm,view}.hpp includes in
  blk_tsr_expr.h with the narrower zip.hpp / any_of.hpp actually used.
• tests/dist_array.cpp, tests/tot_dist_array_part2.cpp: replace 8
  deprecated mktemp(3) calls (-Wdeprecated-declarations on macOS)
  with an mkstemp+close+remove helper that produces a unique filename
  race-free.
• tests/btas.cpp: parenthesize BOOST_REQUIRE_NO_THROW(Tensor(ta_tensor))
  to dodge -Wvexing-parse.

CI — -DTA_WERROR=ON everywhere.

• GHA: 8 matrix cells (macOS/Ubuntu × Pthreads/PaRSEC × Release/Debug).
• GitLab: CUDA-capable runner.

Test plan

• Local Release build (macOS arm64, AppleClang, PaRSEC, no
  ScaLAPACK/CUDA): tiledarray, ta_test, examples-tiledarray
  all build clean with -DTA_WERROR=ON.
• tiledarray-targets.cmake: zero Werror, no
  tiledarray_internal_warnings references.
• -Werror lands on TA src/tests/examples and MADworld; absent
  from parsec/btas/blaspp/lapackpp/umpire/range-v3.
• GHA: all 8 cells green (covers ScaLAPACK on each).
• GitLab: CUDA cell green.
• Spot-check downstream build (MPQC pulling TA via FetchContent)
  does not inherit -Werror.

Blocking dependency

This PR pins MADNESS to the head of m-a-d-n-e-s-s/madness#693 so that
MADNESS_WERROR=ON propagation actually does something. Merge order:
land #693 first, then bump this PR's pin to the merge SHA before
merging here.