fix: memory leak in model.predictImageSet. (intel-analytics#2557)

* fix: memory leak in `model.predictImageSet`.

There're three reasons of memory leak.

1. repeat allocations in bigquant, which will be fixed in BigDL-core.
2. repeat clone module but no release. `model.predictImageSet` will new
   Predictor again and again.
2. share weights.

This patch add a `StorageManager` which contains a concurrent hash map
to maintain all allocations of native memory/resources and prevent
duplicate release. It's also helpful for debug.

* fix: delete .

* refator:  as the API for AbstractModule

* fix: distribute predictor memory leak

* fix: move delete operation to ModelBroadcast

* refinement per review

* fix ut

* fix scala version issue

dl/src/main/scala/com/intel/analytics/bigdl/common/Util.scala

@@ -19,7 +19,7 @@ package com.intel.analytics.bigdl.utils
 import java.io._
 
 import com.intel.analytics.bigdl._
-import com.intel.analytics.bigdl.nn.{Container, Graph}
+import com.intel.analytics.bigdl.nn.Container
 import com.intel.analytics.bigdl.nn.tf.Const
 import com.intel.analytics.bigdl.tensor.TensorNumericMath.{NumericWildcard, TensorNumeric}
 import com.intel.analytics.bigdl.tensor._
@@ -166,6 +166,10 @@ object Util {
     var i = 0
     while (i < tensors.length) {
       if (tensors(i) != null) {
+        if (tensors(i).getTensorType == QuantizedType) {
+          tensors(i).toQuantizedTensor.release()
+        }
+
         tensors(i).set()
       }
       i += 1
@@ -179,10 +183,23 @@ object Util {
     var i = 0
     while (i < localWeightBias.length) {
       if (localWeightBias(i) != null) {
-        localWeightBias(i).set(broadcastWeightBias(i))
+        clearAndSet(localWeightBias(i), broadcastWeightBias(i))
       }
       i += 1
     }
+
+    def clearAndSet(old: Tensor[T], other: Tensor[T]): Unit = {
+      if (old.getTensorType == QuantizedType && other.getTensorType == QuantizedType) {
+        val quantOld = old.asInstanceOf[QuantizedTensor[T]]
+        val quantOther = other.asInstanceOf[QuantizedTensor[T]]
+
+        if (quantOld.getNativeStorage != quantOther.getNativeStorage) {
+          quantOld.release()
+        }
+      }
+
+      old.set(other)
+    }
   }
 
   private[bigdl] def initGradWeightBias[T: ClassTag](