Logs generated before calling LearnerBuilder::build are discarded in some examples

[towerpark]

Describe the bug
In some examples, logs generated before calling LearnerBuilder::build don't get saved to the log file.

Some of the examples affected by this issue are: mnist, text_classification, text_generation. Examples of the discarded logs are of target burn_train::metric::cuda, cubecl_wgpu::runtime, hf_hub, all of which are generated at the pre-model-fitting stage.

The cause is that the built-in logger doesn't get installed until LearnerBuilder::build is called, and no other logger is set before that, so those logs are ignored by the log macros.

To Reproduce
A couple of extra lines will appear at the head of the log file if we set a logger manually as follows instead of relying on LearnerBuilder to do it:

1. Create an instance of FileApplicationLoggerInstaller (from burn-train) and install it at the begin of the program;
2. Call .with_application_logger(None) before building the learner.

Expected behavior
All the logs written via the log crate should be saved into log files.

Maybe LearnerBuilder::tracing_logger should default to None so that users can be freed from making the .with_application_logger(None) call every time they want to use Learner without losing any logs.

====
UPDATE: Added the cause.

[laggui]

Thanks for opening this!

I'll link this comment from our previous discussion to add the initial context.

[towerpark]

Thanks for adding that.
I should have included the link in the post 😅