Fix on log and metric issue on Windows when using pytorch with multithread

[Ivanfangsc]

The problem with log and metric when using torch.utils.data.DataLoader with num_workers>0 on windows is known for a long time. I find that the log and metric file is whipped of its content whenever a DataLoader is enumerated. I reached out to pytorch for this, and they told me it is a issue with the multiprocessing library. It uses on default spawn mode in windows, in which

unnecessary file descriptors and handles from the parent process will not be inherited (https://docs.python.org/3/library/multiprocessing.html#contexts-and-start-methods)

This cause the log and metric file which are opened at top level in __init__.py and runtime/platform/local.py. Thankfully there is a simple workaround on this, that is to change the file to 'a' open mode. I tested it on my local machine and it works.

[ghost]

All CLA requirements met.