add IPO preference optimization

[yuetongz417]

Summary

Adds Identity Preference Optimization (IPO) as a built-in offline preference-training algorithm.

Changes

• Add a sequence-level IPO loss following Eq. 17 of the IPO paper using summed response log-probabilities.
• Reuse the existing DPO preference trainer and reference-model scoring workflow.
• Add IPOTrainerConfig and the --ipo-beta option.
• Register ipo in the algorithm registry.
• Add IPO reference-checkpoint resolution.
• Add CPU tests for the loss, configuration, registry integration, and model-reference resolution.
• Update CLI help and documentation.

The implementation currently uses sequence-summed log-probabilities rather than length-normalized averages.

[xsuler]

@yuetongz417 Hi, it would be better if you add an example in the examples folder and add an evaluation report of this implementation.