wandb/tensorboard loggers set default init to False

[brian-dellabetta]

SUMMARY:
If wandb is installed, an empty logging config will default to including wandb, causing workflows to hang on the following prompt:

wandb: Using wandb-core as the SDK backend.  Please refer to https://wandb.me/wandb-core for more information.
wandb: (1) Create a W&B account
wandb: (2) Use an existing W&B account
wandb: (3) Don't visualize my results
wandb: Enter your choice: 

This resolves by setting default behavior to False, expecting user to provide a config in order to use wandb.
Resolves #976
Resolves #934

TEST PLAN:
No new src code

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[markurtz]

@brian-dellabetta do we not have a better pathway for determining if wandb is configured and in a non-blocking usable state? Since the user doesn't have easy pathways into configuring the loggers currently, especially for our oneshot APIs, I'm a bit cautious of essentially removing metric logging for those completely.

Separately, though, we do need to refactor how we're handling loggers and their general setup, so that can wait til then.

[brian-dellabetta]

@brian-dellabetta do we not have a better pathway for determining if wandb is configured and in a non-blocking usable state? Since the user doesn't have easy pathways into configuring the loggers currently, especially for our oneshot APIs, I'm a bit cautious of essentially removing metric logging for those completely.

Separately, though, we do need to refactor how we're handling loggers and their general setup, so that can wait til then.

@markurtz this part of the codebase is super opaque to me, and because oneshot just takes kwargs it is not clear to me where a user could configure this. Right now this is blocking two users from running anything, so this is a hotfix that should be considered more deeply later on (probably overlapping current work of Kyle and George for their refactoring).

But i think this indicates there is a whole lot of code involved in training that axolotl/llm-foundry/instructlab take care of, and that we're better off connecting into them for more sophisticated workflows that take care of all that