[feat] Dump full configuration in experiment-dir

[RaymondLi0]

🧐 Problem Description

There are lots of parameters in fast-llm. Sometimes we want to double-check what defaults were used in a training run. Currently, we need to dig into the code to check the defaults value.

💡 Proposed Solution

Increase the verbosity of the configuration dumped in experiment-dir (instead of only the core and non-default values).
We could increase to FieldHint.feature or FieldHint.performance for example.

📈 Potential Benefits

Easily inspect all the parameters used for a given run.
Would also make it easier to re-use dumped configs in later runs (default values could change in future versions of fast-llm for example)

[RaymondLi0]

Although I can see why it's convenient to only look at the non-default values.
We could dump multiple configs with different levels of verbosity? Or would it be too much?

[jlamypoirier]

Although I can see why it's convenient to only look at the non-default values. We could dump multiple configs with different levels of verbosity? Or would it be too much?

The problem is there are already hundreds of parameters, and usually most of them are irrelevant, so a full config is an unreadable mess. I'm open to having two separate files though, ex. config.yaml and config_summary.yaml