Fix the race-condition issue on `IgnorableColumns`

Since the Ruby threads are preemptive on the user level, the scheduler
can decide to do the context switching to give the execution to another
thread after the conditional check. If this happens, two threads will
generate the exact same object. We can avoid this by just synchronizing
the whole block.

Since the `Mutex` is not `reentrant` in Ruby, I've also changed the way
we synchronize the critical section with `Monitor`.

app/models/concerns/ignorable_columns.rb

@@ -31,15 +31,13 @@ module IgnorableColumns
     alias_method :ignore_column, :ignore_columns
 
     def ignored_columns_details
-      unless defined?(@ignored_columns_details)
-        IGNORE_COLUMN_MUTEX.synchronize do
-          @ignored_columns_details ||= superclass.try(:ignored_columns_details)&.dup || {}
-        end
-      end
+      return @ignored_columns_details if defined?(@ignored_columns_details)
 
-      @ignored_columns_details
+      IGNORE_COLUMN_MONITOR.synchronize do
+        @ignored_columns_details ||= superclass.try(:ignored_columns_details)&.dup || {}
+      end
     end
 
-    IGNORE_COLUMN_MUTEX = Mutex.new
+    IGNORE_COLUMN_MONITOR = Monitor.new
   end
 end

spec/models/concerns/ignorable_columns_spec.rb

@@ -59,6 +59,14 @@ RSpec.describe IgnorableColumns do
       it_behaves_like 'storing removal information'
     end
 
+    context 'when called on a subclass without setting the ignored columns' do
+      let(:subclass) { Class.new(record_class) }
+
+      it 'does not raise Deadlock error' do
+        expect { subclass.ignored_columns_details }.not_to raise_error
+      end
+    end
+
     it 'defaults to empty Hash' do
       expect(subject.ignored_columns_details).to eq({})
     end