What does this PR do?

This PR tries to improve our DDP backend. Most of the changes tries to improve some corner case of SpeechBrain.

List of additions/modifications:

• add context manager to make sure that a block of code is always entering in __enter__ and __exit__ which makes sure that MAIN_PROC_ONLY is always increased/decreased even when a sub-function is throwing an exception (which in the past could lead to an issue).
• fix an issue with the collect_files(). Indeed, in DDP you expect to see in save/ a symlink of your pre-trained model. This indeed, led to an issue where only the main proc could load the LM while the other processes where throwing an error because they couldn't load the collected models. One effective solution was to turn internal_ddp_handling flag to True but it is not necessary as run_on_main does also work on a system that is not distributed.
• Add a new get_logger function as a drop-in replacement to logger.get_logger. The key idea here is to have the possibility to choose if we want to log on all the processes or only the main proc. Generally speaking, we only want the main proc to log informations (e.g. loading of ckpts etc), in some cases like seed_everything you would like to know the seed for each proc but these are corner case which can be solved by just passing main_process_only=False. This new logger leaves us also the possibility to only emit one warning as wanted in some previous PRs.
• By default, now, we are leveraging Tensor Floats 32 which is absolutely critical and doesn't leads to any instabilities.
• Improvement of DDP barrier function to support the case where we use NCCL.
• Some new distributed functions that can be helpful.

PR review