Register only the LLVM targets this build can emit for, and build the bench on Windows

[dougchansan]

What

main does not link against the official LLVM 20 Windows package.

target.cpp calls InitializeAll*, which expands to every target the host LLVM was configured with. That package ships seven (AArch64 ARM X86 BPF WebAssembly RISCV NVPTX) while CMakeLists.txt linked four, so the link fails with 14 unresolved LLVMInitialize* symbols for BPF, WebAssembly and NVPTX.

Why this shape

Naming X86 and AArch64 unconditionally fixes Windows but breaks against an LLVM built without one of them — the call would not even compile. So both the registration and the linked component list are derived from LLVM_TARGETS_TO_BUILD, which LLVMConfig.cmake reports:

if("X86" IN_LIST LLVM_TARGETS_TO_BUILD)
    list(APPEND DOLRECOMP_LLVM_COMPONENTS X86CodeGen X86AsmParser)
    target_compile_definitions(dr_llvm PRIVATE DOLLLVM_HAVE_X86_TARGET=1)
endif()

with matching #if defined(...) guards in target.cpp, and a FATAL_ERROR if neither target is present.

Also

benchmarks/llvm_backend_bench.c used clock_gettime/CLOCK_MONOTONIC, which MSVC does not provide, so the benchmark target never built on Windows. Adds a QueryPerformanceCounter path; the POSIX path is unchanged.

Verification

• 28/28 ctest on Windows (MSVC 19.50, LLVM 20.1.8 — 7 targets)
• 28/28 ctest on a Raspberry Pi 4 (GCC 14.2, LLVM 19 — 21 targets)

Those two LLVM builds have very different target sets, which is what exercises the gating in both directions.

[siahisaforker]

if i have to review another PR i will explode and die

[dougchansan]

do a flip!