i#3526: Apply i#2089 scheme to client threads.

When creating client threads, we still attempted to invalidate the thread's TLS by writing
a null selector to its segment register. In 64-bit mode, this seemed to work most of the
time because presumably Intel and AMD are then zero'ing out the upper 32-bit of the hidden
segment's base. Intel describes this in Vol. 3A 3.4.4, although not completely clear.

This patch applies the same scheme introduced and described in #2089.

Before this patch, the test code_api|client.thread ran into all kinds of assertions,
because the client thread interfered with the parent's TLS. These failures are now gone.

Fixes #3526
Issue: #2089

core/arch/x86/x86.asm

@@ -1586,8 +1586,6 @@ GLOBAL_LABEL(dynamorio_clone:)
 # endif
         cmp      REG_XAX, 0
         jne      dynamorio_clone_parent
-        /* avoid conflicts w/ parent's TLS by clearing our reg now */
-        mov      SEG_TLS, ax
         pop      REG_XCX
         call     REG_XCX
         /* shouldn't return */

core/unix/os.c

@@ -3719,9 +3719,11 @@ dr_create_client_thread(void (*func)(void *param), void *arg)
 #        endif
     LOG(THREAD, LOG_ALL, 1, "dr_create_client_thread xsp=" PFX " dstack=" PFX "\n", xsp,
         get_clone_record_dstack(crec));
+    os_clone_pre(dcontext);
     thread_id_t newpid =
         dynamorio_clone(flags, xsp, NULL, IF_X86_ELSE(IF_X64_ELSE(NULL, &desc), NULL),
                         NULL, client_thread_run);
+    os_clone_post(dcontext);
     /* i#501 switch to app's tls before creating client thread */
     if (IF_CLIENT_INTERFACE_ELSE(INTERNAL_OPTION(private_loader), false))
         os_switch_lib_tls(dcontext, false /*to dr*/);