[SPARK-14911] [CORE] Fix a potential data race in TaskMemoryManager

## What changes were proposed in this pull request?

[[SPARK-13210][SQL] catch OOM when allocate memory and expand array](37bc203c8d) introduced an `acquiredButNotUsed` field, but it might not be correctly synchronized:
- the write `acquiredButNotUsed += acquired` is guarded by `this` lock (see [here](https://github.com/apache/spark/blame/master/core/src/main/java/org/apache/spark/memory/TaskMemoryManager.java#L271));
- the read `memoryManager.releaseExecutionMemory(acquiredButNotUsed, taskAttemptId, tungstenMemoryMode)` (see [here](https://github.com/apache/spark/blame/master/core/src/main/java/org/apache/spark/memory/TaskMemoryManager.java#L400)) might not be correctly synchronized, and thus might not see `acquiredButNotUsed`'s most recent value.

This patch makes `acquiredButNotUsed` volatile to fix this.

## How was this patch tested?

This should be covered by existing suits.

Author: Liwei Lin <lwlin7@gmail.com>

Closes #12681 from lw-lin/fix-acquiredButNotUsed.

core/src/main/java/org/apache/spark/memory/TaskMemoryManager.java

@@ -114,7 +114,7 @@ public class TaskMemoryManager {
   /**
    * The amount of memory that is acquired but not used.
    */
-  private long acquiredButNotUsed = 0L;
+  private volatile long acquiredButNotUsed = 0L;
 
   /**
    * Construct a new TaskMemoryManager.