ui: save gradient checkpointing option by default

[bghira]

This pull request improves how configuration arguments—especially boolean flags—are handled and saved, ensuring greater clarity and consistency in the training configuration. The main changes focus on making boolean values explicit in argument lists and guaranteeing that certain important fields are always saved, even when defaults would otherwise be omitted.

Improvements to argument handling and configuration saving:

• Boolean arguments are now always passed explicitly with their value (e.g., --flag=true or --flag=false), rather than relying on presence/absence, to avoid ambiguity and ensure that argparse defaults are properly overridden (simpletuner/helpers/configuration/loader.py).

• Introduced an always_save_fields set to specify configuration fields (such as gradient_checkpointing) that must always be saved, regardless of whether their value matches the default, to avoid ambiguity in training runs (simpletuner/simpletuner_sdk/server/services/training_service.py).

• Updated the logic for pruning default values: fields in always_save_fields are now always saved, even when preserve_defaults is enabled, ensuring critical configuration is never omitted (simpletuner/simpletuner_sdk/server/services/training_service.py).