arch: cxd56xx: Fix the pause handler for SMP

Summary:
- I noticed that sched_add_readytorun() runs on multiple CPUs simultaneously
- Finally, I found the root cause which was described in TODO
- Actually, the task newly scheduled on remote CPU did not acquire g_cpu_irqlock
- This commit fixes this issue by adding a critical section to the pause handler
- Which will acquire g_cpu_irqlock on the remote CPU explicitly

Impact:
- SMP only

Testing:
- Tested with spresense:wifi_smp (NCPUS=2 and 4)
- Run smp, ostest, nxplayer

Signed-off-by: Masayuki Ishikawa <Masayuki.Ishikawa@jp.sony.com>

arch/arm/src/cxd56xx/cxd56_cpupause.c

@@ -283,6 +283,7 @@ int up_cpu_paused(int cpu)
 int arm_pause_handler(int irq, void *c, FAR void *arg)
 {
   int cpu = up_cpu_index();
+  int ret = OK;
 
   DPRINTF("cpu%d will be paused \n", cpu);
 
@@ -297,10 +298,21 @@ int arm_pause_handler(int irq, void *c, FAR void *arg)
 
   if (spin_islocked(&g_cpu_paused[cpu]))
     {
-      return up_cpu_paused(cpu);
+      /* NOTE: up_cpu_paused() needs to be executed in a critical section
+       * to ensure that this CPU holds g_cpu_irqlock. However, adding
+       * a critical section in up_cpu_paused() is not a good idea,
+       * because it is also called in enter_critical_section() to break
+       * a deadlock
+       */
+
+      irqstate_t flags = enter_critical_section();
+
+      ret = up_cpu_paused(cpu);
+
+      leave_critical_section(flags);
     }
 
-  return OK;
+  return ret;
 }
 
 /****************************************************************************