fix vendored libdeflate _rotr* macros colliding with mingw stdlib.h

compiler_gcc.h defined _rotl/_rotr/_rotl64/_rotr64 as macros before
mingw's <stdlib.h> got a chance to declare them as functions. Any
translation unit that did `#include "lib_common.h"` (which transitively
pulls in compiler_gcc.h) and later included <stdlib.h> — utils.c does
exactly this — would have the mingw function declarations textually
substituted by the macros, producing a syntax error.

Force <stdlib.h> in from compiler_gcc.h on mingw so the function
declarations are parsed before the macro names get shadowed. The
later #include in utils.c becomes a no-op via the header guard, and
call sites still resolve to the macros below (so the generated code
is unchanged on every other compiler/target).

This unblocks the Crafter.Build CI's x86_64-w64-mingw32 cross-compile,
which has been failing for a while when building Crafter.Asset →
gdeflate → libdeflate as a dep.

lib/gdeflate/libdeflate/common/compiler_gcc.h

@@ -203,7 +203,19 @@ typedef char  __v64qi __attribute__((__vector_size__(64)));
 /*
  * Setup rotation macros similar to MSVS intrinsics.
  * These should recognized by compilers.
+ *
+ * mingw's <stdlib.h> declares _rotl/_rotr/_rotl64/_rotr64 as functions, so
+ * a translation unit that pulls compiler_gcc.h in first (libdeflate's own
+ * utils.c does exactly that) and then includes <stdlib.h> would see the
+ * function declarations mangled by these macros. Force <stdlib.h> in here
+ * on mingw so the declarations get parsed before the macro names get
+ * shadowed; the header guard makes the later #include a no-op and call
+ * sites still resolve to the macros below.
  */
+#if defined(__MINGW32__) || defined(__MINGW64__)
+#  include <stdlib.h>
+#endif
+
 #ifndef _rotr16
 #define _rotr16(x,n)	((x>>n) + (x<<(16-n)))
 #endif