arm/arm64: Rework the cache maintenance in asm_mmu_disable

asm_mmu_disable is overly ambitious and provably incorrect:

1. It tries to clean and invalidate the data caches for the *entire*
memory, which is highly unnecessary, as it's very unlikely that a test
will write to the entire memory, and even more unlikely that a test will
modify the text section of the test image.

2. There is no corresponding dcache invalidate command for the entire
memory in asm_mmu_enable, leaving it up to the test that disabled the
MMU to do the cache maintenance in an asymmetrical fashion: only for
re-enabling the MMU, but not for disabling it.

3. It's missing the DMB SY memory barrier to ensure that the dcache
maintenance is performed after the last store executed in program order
before calling asm_mmu_disable.

Fix all of the issues in one go, by doing the cache maintenance only for
the stack, as that is out of the control of the C code, and add the missing
memory barrier.

The code used to test that mmu_disable works correctly is similar to the
code used to test commit 410b3bf0 ("arm/arm64: Perform dcache clean
+ invalidate after turning MMU off"), with extra cache maintenance
added:

+#include <alloc_page.h>
+#include <asm/cacheflush.h>
+#include <asm/mmu.h>
 int main(int argc, char **argv)
 {
+       int *x = alloc_page();
+       bool pass = true;
+       int i;
+
+       for  (i = 0; i < 1000000; i++) {
+               *x = 0x42;
+               dcache_clean_addr_poc((unsigned long)x);
+               mmu_disable();
+               if (*x != 0x42) {
+                       pass = false;
+                       break;
+               }
+               *x = 0x50;
+               /* Needed for the invalidation only. */
+               dcache_clean_inval_addr_poc((unsigned long)x);
+               mmu_enable(current_thread_info()->pgtable);
+               if (*x != 0x50) {
+                       pass = false;
+                       break;
+               }
+       }
+       report(pass, "MMU disable cache maintenance");

Signed-off-by: Alexandru Elisei <alexandru.elisei@arm.com>