Introduce multiprocess logger

[muellerzr]

Introduce Multi-Process aware Logger

What does this add?

Adds a new LoggerAdapter and a get_logger helper to help aid with printing logs repeatedly when in multiprocess

Why is it needed?

Users were reporting that the logs can confuse them with the actions actually occuring (such as logging multiple times a model was saved even though it only happened once).

Rather than an onslaught of if statements, this PR takes the simpler approach of creating an Adapter that will do the check for us.

By default it will only log to the main processor, but if all processes should log just pass main_process_only=False (see examples below)

What parts of the API does this impact?

User-facing:

Loggers in Accelerate now accept a main_process_only kwarg

Basic Usage Example(s):

from accelerate.logging import get_logger
logger = get_logger("my_logger")

# If we only want it called in the main process
logger.info("My message!")
# or
logger.info("My message!", main_process_only=True)

# If it should be called across all processes
logger.info("My message!", main_process_only=False)

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Looks really neat!

[sgugger]

Nice!