[clang] target attributes with `arch=` doesn't work if `-march` is set

[usamoi]

This compiles compiles without any flags, but fails with -march=x86-64-v2.

__attribute__((__always_inline__, target("sse"))) void g() {}
__attribute__((target("arch=x86-64,sse"))) void f() { g(); }

clang f.c -shared -march=x86-64-v2
f.c:2:55: error: always_inline function 'g' requires target feature 'sse3', but would be inlined into function 'f' that is compiled without support for 'sse3'
    2 | __attribute__((target("arch=x86-64,sse"))) void f() { g(); }
      |                                                       ^
1 error generated.

This compiles compiles without any flags or with -march=x86-64, -march=x86-64-v2, -march=x86-64-v3, -march=x86-64-v4, but fails with -march=znver1, -march=znver2, -march=znver3, -march=znver4, -march=znver5. It also compiles with -march=native even if sys::getHostCPUName() detects my cpu as znver5.

__attribute__((__always_inline__, target("avx512fp16"))) void g() {}
__attribute__((target("arch=x86-64-v4,avx512fp16"))) void f() { g(); }

clang f.c -shared -march=znver5
f.c:2:65: error: always_inline function 'g' requires target feature 'gfni', but would be inlined into function 'f' that is compiled without support for 'gfni'
    2 | __attribute__((target("arch=x86-64-v4,avx512fp16"))) void f() { g(); }
      |                                                                 ^
1 error generated.

cc tensorchord/pgvecto.rs#650