Don't block during query optimization

In Spark 2.4.0 the CacheManager was updated so it will not execute jobs while it holds a lock.This was introduced in SPARK-23880.

The CacheManager still holds a write lock during the execution of the query optimizer.  For complex queries the optimizer can run for a long time (we see 10-15 minutes for some exceptionally large queries).  This allows only 1 thread to optimize at once.