build(vk): link the Vulkan loader on Windows under -static

[Kenneth-Javier]

Summary

make colibri.exe VK=1 fails at link time under MSYS2 UCRT64 on current dev (4e28e39):

ld.exe: cannot find -lvulkan: No such file or directory
ld.exe: have you installed the static version of the vulkan library ?

Windows has no libvulkan. The loader is vulkan-1.dll, installed by the GPU driver, and MSYS2's mingw-w64-ucrt-x86_64-vulkan-loader ships only its import library, libvulkan-1.dll.a. The Windows LDFLAGS carry -static, and under it ld does not consider .dll.a import libraries for -l at all, so -lvulkan-1 fails the same way (and -Wl,-Bdynamic -lvulkan finds nothing: the library is named vulkan-1). CI builds VK=1 only on Linux (the vulkan job), which is why nothing reported it.

Fix. On Windows, VK=1 appends -Wl,-Bdynamic -lvulkan-1 -Wl,-Bstatic: dynamic lookup for the loader alone, then back to static, so libgomp, winpthreads and libgcc stay linked in as before. Linux keeps -lvulkan. backend_vulkan.o and the SPIR-V rules are unchanged. Resolving the import library's path with $(CC) -print-file-name=libvulkan-1.dll.a also links, but it puts an absolute toolchain path into LDFLAGS and .build-config and costs a parse-time shell call.

Test. tests/test_makefile_platform.py pins the tail of the VK=1 link line from a dry run, for x86_64-w64-mingw32 and x86_64-unknown-linux-gnu, so it runs on any host with make. Against the previous Makefile the mingw case fails ('-lpsapi', '-lvulkan') and the Linux case passes.

Docs. docs/vulkan.md only showed Linux. It gains a short Windows section: the three MSYS2 packages (vulkan-headers, vulkan-loader, and shaderc for glslc), the one added import, the fact that on Windows the shaders are found from the current directory or COLI_VK_SHADERS (the next-to-the-binary lookup in vk_resolve_spv is __linux__ only), and the CI's config.json init probe as a PowerShell snippet, for checking a driver before downloading a model.

Validation

Windows 11, MSYS2 UCRT64 gcc 16.1.0, make 4.4.1, vulkan-loader 1.4.357.0, shaderc 2026.3, Radeon 8060S:

• make colibri.exe VK=1 links, with 0 compiler warnings. Its import table differs from the default build's by exactly one entry, vulkan-1.dll (13 vs 12).
• The docs snippet, run from c\ with PATH set to System32 and Windows only, prints [VK] ready: AMD Radeon(TM) 8060S Graphics and expert tier active, in pwsh 7.6 and in Windows PowerShell 5.1. From another directory it reports cannot open shaders/qmatmul.spv; with COLI_VK_SHADERS set it initialises again.
• Without VK=1 nothing changes: the default colibri.exe command and all 256 test-c commands from make -Bn are identical to dev's, so the default binary and the oracle are unaffected.
• make check (default build): exit 0; 170 C test binaries with no failure, 1570 Python tests OK (192 skipped).
• make -k test-c VK=1: every link finds the loader, and 166 of 170 test binaries build and pass. The other four (test_xdna_qt_state, test_xdna_failure, test_logprob_status, test_ablate_mode) include colibri.c without linking $(VK_OBJ) and stop on undefined coli_vk_*. That gap is separate and exists on Linux too; I'll send it as a follow-up to tests: link backend_vulkan.o into the engine-including tests under VK=1 #1728. With that follow-up on top, all 170 build and pass.
• With CUDA_DLL=1 or HIP_DLL=1 added: all 172 VK=1 link lines carry -lvulkan-1 and none fails to find it. The 52 targets that fail also fail without VK=1: they do not link backend_loader.o.

WSL Ubuntu 24.04, gcc 13.3.0, make 4.3, libvulkan-dev 1.3.275.0, glslc 2023.8, Mesa 25.2.8:

• The VK=1 compile-and-link command from make -Bn is byte-identical to the previous Makefile's, so the Linux build and its diagnostics are unchanged.
• make colibri VK=1 links, ldd shows libvulkan.so.1, and the ci.yml Lavapipe probe prints [VK] ready: llvmpipe and expert tier active.
• The platform tests pass (4/4).

Not verified: the MINGW64 and CLANG64 environments, the LunarG SDK's vulkan-1.lib, Cygwin, cross-compiling with a mingw gcc from Linux, any GPU other than the Radeon 8060S and Lavapipe, and a Windows VK=1 run on real weights (the probe stops after the backend initialises). This PR adds no Windows VK=1 CI job.

• make -C c check
• CUDA changes were tested with make -C c cuda-test (not applicable)
• Performance claims include hardware, commands, and repeatable measurements (no performance claim)
• Performance claims include a validated experiment manifest with raw evidence (no performance claim)

Compatibility

• The default CPU build remains dependency-free (only VK=1 links change, and only on Windows)
• No model files, generated binaries, or benchmark artifacts are included

Checked locally: this merges cleanly with dev, with #1758, and with #1338, which also edits docs/vulkan.md.