NavyThread: drain and destruct ScopedEventBaseThread first

Summary:
As discovered in the unit test https://fburl.com/testinfra/87ndzjg8, it seems that the segfault
could occur while destructing the worker thread, i.e., NavyThread of RegionManager if there are some
tasks are outstanding while accessing the mutex which is another member of NavyThread. This change
makes sure the ScopedEventBaseThread is destructed after draining any pending fibers before any
other member variables.

Reviewed By: therealgymmy

Differential Revision: D49707371

fbshipit-source-id: 3fcac297d759b70a061ceaee209bfac70f962b4b

Author: Jaesoo Lee

cachelib/navy/common/NavyThread.h

@@ -41,10 +41,12 @@ class NavyThread {
   /**
    * Initializes with current EventBaseManager and passed-in thread name.
    */
-  explicit NavyThread(folly::StringPiece name)
-      : th_(name), fm_(&folly::fibers::getFiberManager(*th_.getEventBase())) {}
+  explicit NavyThread(folly::StringPiece name) {
+    th_ = std::make_unique<folly::ScopedEventBaseThread>(name.str());
+    fm_ = &folly::fibers::getFiberManager(*th_->getEventBase());
+  }
 
-  ~NavyThread() = default;
+  ~NavyThread() { th_.reset(); }
 
   /**
    * Add the passed-in task to the FiberManager.
@@ -71,7 +73,7 @@ class NavyThread {
   NavyThread& operator=(const NavyThread& other) = delete;
 
   // Actual worker thread running EventBase and FiberManager loop
-  folly::ScopedEventBaseThread th_;
+  std::unique_ptr<folly::ScopedEventBaseThread> th_;
 
   // FiberManager which are driven by the thread
   folly::fibers::FiberManager* fm_;