Fix a couple of Executor bugs

[huzecong]

1. File handling Now in Executor, files are only opened in train and test and closed afterwards. Originally files were opened at the point of construction, and closed only when the program exits, using an atexit hook. This keeps the Executor object alive and creates a memory leak.

   Here we keep the change, but handles the logic more carefully. A few corner cases were also considered, and we also open files in write_log. This allows writing logs after construction and before training, while the original implementation simply does nothing, which is confusing for users.

2. Metrics not evaluated When Executor uses a SimpleMetric that is not logged, the metric is simply not evaluated. This is problematic for metrics with side effects, e.g. FileWriterMetric used in the BERT examples. We now explicitly evaluate all metrics at the end of each epoch, and the end of validation and testing.

[codecov[bot]]

Codecov Report

Merging #320 into master will increase coverage by 0.14%. The diff coverage is 87.64%.

@@            Coverage Diff             @@
##           master     #320      +/-   ##
==========================================
+ Coverage   79.93%   80.08%   +0.14%     
==========================================
  Files         133      134       +1     
  Lines       11143    11195      +52     
==========================================
+ Hits         8907     8965      +58     
+ Misses       2236     2230       -6     

Impacted Files	Coverage Δ
texar/torch/run/executor.py	79.38% <82.75%> (+0.29%)	:arrow_up:
texar/torch/run/metric/base_metric.py	88.50% <83.33%> (+11.12%)	:arrow_up:
texar/torch/run/metric/__init__.py	100.00% <100.00%> (ø)
texar/torch/run/metric/io.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update a2f1186...6cd0a31. Read the comment docs.