Windows DLL build (RX_SHARED on Windows)

[Force67]

Makes the component build (one DLL per module) work on Windows. Needs Force67/equilibrium#27 (submodule bumped), Force67/libultragui#15 and Force67/zetanet#7, all merged.

PE gives every DLL that links a static library its own copy of its globals, where ELF shares them. So under RX_SHARED on Windows:

• equilibrium base is a DLL (EQ_BASE_SHARED): one Option/Feature registry, log, allocator
• libultragui is a DLL (ULTRAGUI_SHARED): one UI context and component-id table
• SDL3 is a DLL: one event queue (input_touch_sdl_test caught two)
• mimalloc is off: an executable's operator new replacement does not reach DLLs

Cross build tooling:

• nix develop .#mingw: ucrt64 mingw gcc, mcfgthread, OpenSSL for Windows; the toolchain file reads them from the environment instead of hardcoded store paths (the old ones had been collected)
• toolchain: compiler checks link, packages resolve against the mingw sysroot only (it used the host's Linux SDL3), -mbig-obj
• fixes it turned up: path_motion_test included what it used only under RX_HAS_NRD; http_test's local server is POSIX-only

Verified (mingw cross, under Wine, Vulkan on the real GPU):

• full engine builds as 43 DLLs; static and DLL builds agree on 5 scenes + 11 demos (rmse <= 0.0007) and on the test suite
• editor runs; an option set from the environment reaches the render DLL (wireframe capture identical to static)
• Linux static and shared: ctest 155/155 under vkrun, shared editor runs

Not verified: MSVC. Known Windows-port issues, same in static and DLL: Wine's d3d12.dll fails some compute pipelines (the *_d3d12 tests, and the editor, which defaults to D3D12 on Windows), scene_authoring_test writes an unescaped Windows path, world_bake_test runs rxworld by its Linux path, and under Wine a process is occasionally SIGKILLed at exit after a clean shutdown.