CI: Enable unity build for MRVoxels

[oitel]

No description provided.

[Fedr]

Build-step timings: this PR's 2 CI runs vs 4 master runs at the PR's merge base (12e6449, 59379aa, 5ec776c, 658f6ca). I only compared jobs that ran on the same CPU model (from the cpu_model line in each job log). The hosted x64 runners mix at least 6 CPU models, so x64 runs on different CPUs can't be compared directly. All arm64 runners are Cobalt 100.

Job (CPU)	this PR	master	Change
ubuntu22 arm64, g++-12 (Cobalt 100)	8.5, 8.8 m	9.5 – 10.5 m	−14%
ubuntu24 arm64, g++-13 (Cobalt 100)	9.2, 9.0	10.2 – 10.7	−13%
ubuntu26 arm64, g++-15 (Cobalt 100)	9.4, 9.4	9.6 – 10.5	−8%
ubuntu26 arm64, g++-16 (Cobalt 100)	10.0, 9.7	11.0 – 11.4	−12%
linux-vcpkg Debug arm64, GCC 11 (Cobalt 100)	6.8, 6.6	7.2 – 8.4	−12%
linux-vcpkg Release arm64, Clang (Cobalt 100)	4.3, 4.0	4.6 – 4.8	−11%
emscripten Singlethreaded (Cobalt 100)	3.8, 3.7	4.1 – 4.5	−13%
emscripten Multithreaded (Cobalt 100)	3.9, 4.0	4.2 – 4.4	−7%
ubuntu26 x64, g++-16 (EPYC 7763)	14.9, 15.3	17.5, 17.8	−15%
Windows CMake arm64 (Cobalt 100)	13.3, 14.0	13.3 – 14.4	−2.5%
Windows MSBuild arm64 (Cobalt 100), no unity build: control	12.7, 13.5	12.5 – 13.7	−1%

On Linux and Emscripten this saves about 1 – 1.5 minutes per job. In almost every row, the slower PR run beats the fastest master run. The MSBuild control doesn't change, so it isn't a general runner difference.

On MSVC the gain is smaller. Locally (CMake + Ninja, Release, shared PCH), MRVoxels' summed compile time dropped by 39% with 4 batches, but its own wall-clock only by ~10%. The limit is MRFloatGrid.cpp alone, which takes ~38 s.

Suggestion: also enable unity build for MRVoxels in MRVoxels.vcxproj with <EnableUnitySupport>true</EnableUnitySupport>, as MeshLibC2.vcxproj already does for the C bindings. The MSBuild jobs are among the longest Windows jobs (x64 ~17 – 18 min). One trade-off: MSBuild is currently the only CI build of MRVoxels without unity build, so it's the only one that catches a missing #include that another file in the same batch hides.