Revert ZeroGC: stop depending on MethodTable internal layout for IsLargeObject (#3298)

[kkokosa]

Summary

Reverts #3298 ("ZeroGC: stop depending on MethodTable internal layout for IsLargeObject"), restoring the original pObj->GetGCSafeMethodTable()->GetBaseSize() >= LARGE_OBJECT_SIZE implementation of IsLargeObject.

Why

Per review discussion on #3298: the GC/EE data-layout contract is versioned in a structured way, not just by convention.

• LARGE_OBJECT_SIZE (85,000 bytes) is a public constant from the versioned gcinterface.h - safe to depend on regardless.
• The one concrete historical layout break ZeroGC: stop depending on MethodTable internal layout for IsLargeObject #3298 was defending against - MethodTable::Collectible()'s flag bit remap (Converge Representations between NativeAOT and CoreCLR runtime#91821) - bumped EE_INTERFACE_MAJOR_VERSION 1→2 in that same PR. gcenv.object.h's existing shim (g_oldMethodTableFlags), which ZeroGC's dllmain.cpp already plugs into (g_runtimeSupportedVersion.MajorVersion < 2), branches on exactly that signal. So this kind of layout drift is covered by a real, numbered contract, not just convention.
• MethodTable::GetBaseSize() itself has never needed such a compatibility shim.
• ZeroGC can't avoid depending on MethodTable/Object layout anyway - those types are mandatory for the IGCHeap interface signatures it implements. Removing this one call site didn't reduce ZeroGC's actual exposure to layout drift.

Net effect of #3298 was added ongoing cost (a std::map + critical section on every LOH-sized allocation) without closing off a real risk. This PR reverts it back to the simpler, original implementation.

Testing

• Builds clean via src/ZeroGC/native/build.ps1.
• Diff against the commit immediately before ZeroGC: stop depending on MethodTable internal layout for IsLargeObject #3298 for ZeroGCHeap.cpp is empty (exact revert, no residual changes).