add torch-ffn and unsloth-ffn gradient checkpointing modes for speed-up

[bghira]

This pull request expands and clarifies the documentation for the --gradient_checkpointing_backend option across all supported languages, and introduces new experimental guides explaining the tradeoffs and usage of Unsloth-style checkpointing. The main changes include adding new backend options, updating descriptions to reflect these, and providing detailed, language-specific guides for advanced checkpointing strategies.

Documentation improvements:

• The --gradient_checkpointing_backend option now documents four choices: torch, torch-ffn, unsloth, and unsloth-ffn, with clear explanations of each backend's behavior, requirements, and supported model scopes. [1] [2] [3] [4] [5] [6]

Experimental guides:

• Added detailed experimental documentation for Unsloth-style checkpointing in Spanish, Hindi, and Japanese, covering practical usage advice, memory and speed tradeoffs, and best practices for choosing checkpointing strategies. [1] [2] [3]

Internationalization:

• All changes to the --gradient_checkpointing_backend option and the new guides are provided in English, Spanish, Hindi, Japanese, Portuguese, and Chinese, ensuring consistent and clear guidance for users in all supported languages. [1] [2] [3] [4] [5] [6] [7] [8] [9]

These updates ensure that users understand the available checkpointing backends, their tradeoffs, and how to apply them effectively in different training scenarios.