fix(test): make test_stops build on Windows (mkdtemp compat shim)

[woolcoxm]

What

make test halted on the Windows job: tests/test_stops.c uses POSIX mkdtemp(), which MinGW-w64 doesn't declare, so the test failed to compile on Windows (and only there).

tests/test_stops.c:73:9: error: implicit declaration of function 'mkdtemp';
   did you mean 'mktemp'? [-Wimplicit-function-declaration]
make: *** [Makefile:318: tests/test_stops.exe] Error 1

This blocked the entire C-test phase on Windows. (test_uring is correctly gated to Linux via ifneq (,$(LINUX)), so test_stops was the only Windows blocker.)

Fix

Added a compat_mkdtemp shim to c/compat.h, following the file's stated convention — "every platform difference lives here; the .c files stay clean." POSIX mkdtemp takes a char[] template ending in XXXXXX, replaces the X's randomly, mkdirs 0700, and returns the pointer (or NULL). On Windows:

static inline char *compat_mkdtemp(char *tmpl){
    if(!tmpl) return NULL;
    if(!_mktemp(tmpl)) return NULL;       /* fills the trailing X's in place */
    if(_mkdir(tmpl) != 0) return NULL;    /* EEXIST is impossible post-_mktemp */
    return tmpl;
}
#define mkdtemp(tmpl) compat_mkdtemp(tmpl)

_mktemp (CRT, <io.h>) has the same XXXXXX contract; _mkdir needed <direct.h>, added. On Linux compat.h is a complete no-op — POSIX mkdtemp is untouched.

Why a shim and not a test-local fix

The earlier fix in test_stops.c (commit on this file) already moved off /tmp/... to a CWD-relative path to fix the Windows job, but mkdtemp itself was still POSIX-only. A compat.h shim is the project's established pattern for exactly this (see getline, setenv, posix_fadvise, pread, rename...) and keeps the test portable without #ifdef _WIN32 inside the test source.

Verification

On MinGW:

make tests/test_stops.exe   # builds clean (the mkdtemp error is gone)
./tests/test_stops.exe      # all 5 sub-cases pass

The two unrelated -Wall warnings that still show in this branch's build (compat.h:241 pragma, glm.c:1210 g_numa_nodes) are addressed by #350 — this PR is based on origin/dev and is independent of it.