Fix data race and thread leak in ThreadedIter

[hcho3]

Fixes #550

• Prevent "thread leak": Wrap std::thread* with std::unique_ptr<> and ScopedThread so that we can automatically join() a thread whenever an exception is thrown. Do not use raw pointer to store the thread, as it's not exception-safe.

• Eliminate data races in producer signals: Make producer_sig_processed_ and producer_sig_ atomic because 1) Multiple threads access these variables; 2) The variables are often used without a lock.

  Furthermore, we use .store(*, std::memory_order_release) and .load(std::memory_order_acquire) so that a write from one thread is guaranteed to be visible from another thread. See Release-Acquire Ordering.

• Enable building dmlc-core with Sanitizer. For example, to use Thread Sanitizer, run:

  cmake .. -DGOOGLE_TEST=ON -DDMLC_USE_SANITIZER=ON \
       -DDMLC_ENABLED_SANITIZERS="thread" -DCMAKE_BUILD_TYPE=Debug ..

cc @larroy @jermainewang

[jermainewang]

Consider enable sanitizer in the CI after these fixes?

[hcho3]

@jermainewang There's still a long way to go until we resolve all errors reported by the Sanitizer: #553

[trivialfis]

We can fix them one by one?

[hcho3]

@trivialfis Yes, I plan to merge this PR pretty soon (after addressing comments from @junrushao1994), and then tackle remaining concurrency issues in subsequent PRs.

[hcho3]

@trivialfis Since there are outstanding concurrency issues, I'm afraid we are not yet able to add sanitizer to CI.

[hcho3]

@trivialfis For now, I added the sanitizer test and marked it non-blocking.