Fixed cuda re-initialization issue when multiprocessing

[bohJiang12]

Issue description

Ran the app doctr-wrapper on both standalone and containerized ways, and faced error:

Cannot re-initialize CUDA in forked subprocess. To use CUDA with multiprocessing, you must use the 'spawn' start method

Implementations

• Imported torch.multiprocessing module
• Forced the start method of multiprocess as spawn instead of fork, such that multiprocessing.set_start_method('spawn', force=True)
• Then placed the setting outside of the class

[keighrim]

@bohJiang12 found that the error doesn't occur when the app runs in development mode (i.e., w/o --production flag). Hence this might be an issue in the SDK code, as also discussed in pytorch/pytorch#68861. I will investigate the gunicorn wrapper in the SDK code.

[keighrim]

Just tried to run this fix, and got the same errors. Will investigate further.

[keighrim]

The root cause of this issue was the previous implementation's use on Python's built-in concurrent for multithreading, which doesn't work with CUDA. Given that the application has been single-threaded as of v1.2, this specific error related to CUDA re-initialization should be resolved. Closing this pull request without merging the changes.