Refactor KTO coordinated with DPO [b/N]: Simplify truncation logic.

This PR simplifies the truncation logic by removing max_prompt_length and truncation_mode parameters, keeping only max_length.

This simplification makes the KTO trainer significantly easier to use and understand, while maintaining correct truncation behavior.

Follow-up to:

• Refactor KTO coordinated with DPO [a/N]: Remove encoder-decoder support #4792

Part of:

• KTO refactoring #4786 (comment)

Coordinated with DPO refactoring, as discussed with @qgallouedec :

• Refactor DPO #3906

Key Changes

1. KTOConfig

   • Removed max_prompt_length parameter and documentation
   • Removed truncation_mode parameter and documentation
   • Now only uses max_length for sequence length control

2. KTOTrainer

   Simplified _process_tokens() function:

   • Removed complex two-stage truncation logic
   • Old logic: First truncate prompt (with keep_start/keep_end modes), then truncate completion
   • New logic: Simple single-stage truncation from the end (completion only)
   • Updated docstring to reflect simpler behavior

   Trainer Initialization:

   • Removed max_prompt_length setup logic
   • Removed self.max_prompt_length attribute
   • Removed self.truncation_mode attribute
   • Removed both parameters from fn_kwargs passed to _process_tokens()

3. Tests

   • Updated test to remove truncation_mode and max_prompt_length from fn_kwargs