unitsync: link generic cpu_topology instead of per-platform detection

[tomjn]

What

unitsync links a generic, platform-agnostic cpu_topology implementation instead of the per-platform Platform/{Linux,Win,Mac}/CpuTopology.cpp, removing unitsync's dependency on real CPU topology detection on all platforms.

This is an alternative to #3042's "add Mac/CpuTopology.cpp to unitsync" — it instead makes unitsync stop needing per-platform topology at all, which is what @sprunk asked for on that PR.

Why

unitsync wraps FileSystemInitializer::Initialize() in the thread pool so archive scanning/hashing parallelizes via for_mt, so it genuinely needs the pool. It does not pin threads, so it does not need real CPU topology (P/E core masks, cache groups, pin policy).

The dependency was incidental: CPUID::EnumerateCores() (called from the CPUID singleton ctor) eagerly calls cpu_topology::GetProcessorMasks() + GetProcessorCache(), and Threading::GetChosenThreadPinPolicy() calls cpu_topology::GetThreadPinPolicy(). Those are link-time references in TUs unitsync compiles, so it had to link a per-platform CpuTopology.cpp even though it never pins.

Approach

A generic CpuTopologyGeneric.cpp reports every logical core as a single group of performance cores (so CPUID::EnumerateCores() still derives a correct logical-core count for the pool) and requests THREAD_PIN_POLICY_NONE. unitsync links that instead of the per-platform file. The change is contained to tools/unitsync/CMakeLists.txt + the new file; no shared engine CMake or threading logic is touched, so the engine/dedicated server keep linking the real per-platform topology for sim-worker pinning.

Note on the "make CPUID lazy" alternative

An earlier idea was to make CPUID::EnumerateCores() compute topology lazily so the core-count path wouldn't pull cpu_topology. That does not work for unitsync: it's a SHARED library and on macOS/Windows the build sets no -fvisibility=hidden (gated to Linux in TestCXXFlags.cmake), so the Threading/CpuTopologyCommon free functions that reference the symbols stay exported and can't be dead-stripped. Deferring when a call runs doesn't remove a link-time reference. So a generic impl is the approach that actually drops the per-platform file.

This satisfies the dependency with a trivial stub rather than eliminating the symbol references entirely; a literal "drop" would require #ifdef UNITSYNC stubs through CpuID.cpp/Threading.cpp/CpuTopologyCommon.cpp (shared core files). Happy to go that route instead if preferred.

Verification

Built unitsync on macOS (Homebrew GCC 16):

• Links cleanly.
• Link line contains CpuTopologyGeneric.cpp.o, not Mac/CpuTopology.cpp.o.
• nm on libunitsync.dylib: zero undefined cpu_topology symbols; none of the real platform internals (detect_cpu_vendor, collect_intel, get_thread_siblings) are linked.

Engine targets are untouched in their source wiring, so they keep the real topology.

Refs #3042.

Assisted by Claude Code; verified by building on macOS.

[sprunk]

Turns out the unitsync use of threading is reasonable so I think the original approach in that other PR (to actually add the proper mac cpu topology) is better after all.

[tomjn]

Closed in favor of #3042