gfx_thumbnail_status_atomic: make the documented STRESS_ITERS override usable

The file tells you the iteration count is "Overridable
(-DSTRESS_ITERS=...) for constrained environments", but the Makefile
gives you nowhere to put it.  The only lever is CFLAGS=, and a
command-line CFLAGS= replaces every += the Makefile does rather than
adding to it -- so following the advice silently drops -DHAVE_THREADS,
and the run reports

    [pass] STORE/LOAD round-trip on all four status values
    [skip] HAVE_THREADS not defined; static checks only
    ALL OK

exiting 0 with the entire stress section switched off.  The skip line
is there, to the test's credit, but "ALL OK" and a zero exit are what a
CI job looks at.

Add an EXTRA_CFLAGS hook and point the comment at it.  The measured
costs go in the comment too, since the reason to reach for the knob is
always the same:

    20000    2.9 s
    100000   14.4 s
    250000   36 s
    1000000  ~145 s   (the default)

all on a single core.

Found while running every sample in samples/ end to end after the
vulkan validation one turned out to be passing without reaching a
device.  That was a test that checked nothing; this is a documented
escape hatch that turns the test off if you use it as documented.  Same
family.

samples/gfx/gfx_thumbnail_status_atomic/Makefile

@@ -29,6 +29,17 @@ ifeq ($(HAVE_THREADS),1)
    LDFLAGS += -lrt
 endif
 
+# Extra flags for the caller, e.g. the -DSTRESS_ITERS= override the
+# test documents:
+#
+#     make EXTRA_CFLAGS=-DSTRESS_ITERS=20000
+#
+# Passing those through CFLAGS= instead would clobber everything set
+# above, -DHAVE_THREADS included, and the test would quietly report
+# "[skip] HAVE_THREADS not defined; static checks only" while still
+# exiting 0 -- a pass with the whole stress section switched off.
+CFLAGS += $(EXTRA_CFLAGS)
+
 OBJS := $(SOURCES:.c=.o)
 
 ifneq ($(SANITIZER),)

samples/gfx/gfx_thumbnail_status_atomic/gfx_thumbnail_status_atomic_test.c

@@ -214,10 +214,21 @@ typedef char _gfx_thumb_status_size_check[
  * out a missing-barrier bug under TSan within a few seconds; on
  * real weak-memory hardware (qemu-aarch64) it surfaces faster.
  * Adjust upward if the bug becomes harder to reproduce.
- * Overridable (-DSTRESS_ITERS=...) for constrained environments:
- * the lock-step handshake costs scheduler timeslices per
- * iteration on a single-CPU host, and TSan's happens-before
- * detection does not depend on the count. */
+ * Overridable for constrained environments: the lock-step
+ * handshake costs scheduler timeslices per iteration on a
+ * single-CPU host, and TSan's happens-before detection does not
+ * depend on the count.  Pass it through the Makefile's
+ * EXTRA_CFLAGS, not CFLAGS:
+ *
+ *     make EXTRA_CFLAGS=-DSTRESS_ITERS=20000
+ *
+ * CFLAGS= on the command line replaces everything the Makefile
+ * sets, -DHAVE_THREADS included, and the run then reports
+ * "[skip] HAVE_THREADS not defined" and exits 0 with the whole
+ * stress section switched off.
+ *
+ * Measured on one core: 2.9 s at 20000, 14.4 s at 100000, 36 s
+ * at 250000, ~145 s at the default. */
 #ifndef STRESS_ITERS
 #define STRESS_ITERS 1000000
 #endif