Switch to USE_COMPILER_TLS on Cosmo

Issue #666 (bdwgc).

In thread-local allocation implementation, use "__thread" keyword to
declare GC_thread_key and directly access the key instead of use of
pthread_set/getspecific.

* include/private/thread_local_alloc.h [!USE_PTHREAD_SPECIFIC
&& !USE_WIN32_SPECIFIC && !USE_WIN32_COMPILER_TLS && !USE_COMPILER_TLS
&& !USE_CUSTOM_SPECIFIC && !GC_WIN32_THREADS && !HOST_ANDROID
&& !(LINUX && GC_GNUC_PREREQ(3,3)) && COSMO && (GC_GNUC_PREREQ(4,4)
|| GC_CLANG_PREREQ(3,9))] (USE_COMPILER_TLS): Define.

include/private/thread_local_alloc.h

@@ -73,7 +73,7 @@ EXTERN_C_BEGIN
 #        define USE_COMPILER_TLS
 #      endif
 
-#    elif (defined(FREEBSD)                                                   \
+#    elif (defined(COSMO) || defined(FREEBSD)                                 \
            || (defined(NETBSD) && __NetBSD_Version__ >= 600000000 /* 6.0 */)) \
         && (GC_GNUC_PREREQ(4, 4) || GC_CLANG_PREREQ(3, 9))
 #      define USE_COMPILER_TLS