Add try/finally block to yield in reporter.py

[keisukefukuda]

background

In reporter.py, a logging context stack is maintained as its internal thread-local variable named _thread_local.reporters.

In Reporter.scope() function, a Reporter object is pushed and popped to the stack in __enter__() and __exit__() function.

the bug

If the yield omits an exception, the __exit__() is not called and the stack becomes an inconsistent state. In most use cases, such runtime error causes program termination thus the bug does not matter a lot.

However, in rare cases, such as #8384, it matters.
The bug forces the whole pytest process to terminate during a test.

fix

This PR fixes the problem by adding try/finally block around the yield sentence so the pop operation is done correctly and the stack is kept consistent.

Appendix:

The error message from #8384 test looks like this:

tests/chainermn_tests/links_tests/test_batch_normalization.py .F

=================================== FAILURES ===================================
____________ test_multi_node_bn_cpu[NaiveCommunicator-mpi-float16] _____________

communicator_class = <class 'chainermn.communicators.naive_communicator.NaiveCommunicator'>
backend = 'mpi', dtype = <class 'numpy.float16'>

    @pytest.mark.parametrize(('communicator_class', 'backend', 'dtype'), [
        (NaiveCommunicator, 'mpi', numpy.float16),
        (NaiveCommunicator, 'mpi', numpy.float32),
        (NaiveCommunicator, 'mpi', chainer.mixed16)])
    def test_multi_node_bn_cpu(communicator_class, backend, dtype):
        comm = create_communicator(communicator_class, mpi_comm,
                                   use_gpu=False)
>       check_multi_node_bn(comm, backend=backend, dtype=dtype)

tests/chainermn_tests/links_tests/test_batch_normalization.py:240:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
tests/chainermn_tests/links_tests/test_batch_normalization.py:137: in check_multi_node_bn
    l1 = m1(x, y)
chainer/link.py:287: in __call__
    out = forward(*args, **kwargs)
chainer/links/model/classifier.py:145: in forward
    reporter.report({'loss': self.loss}, self)
chainer/reporter.py:246: in report
    current.report(values, observer)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

[emcastillo]

Jenkins, test this please

[chainer-ci]

Jenkins CI test (for commit e5e6595, target branch master) succeeded!

[chainer-ci]

@emcastillo This pull-request is marked as to-be-backported, but the corresponding backport PR could not be found. Could you check?