Add UniLora tuner to PEFT

[KaiyangLi1992]

Motivation

This PR adds UniLora, a LoRA-style parameter-efficient fine-tuning method that reconstructs LoRA parameters from a shared low-dimensional vector.

This is a fresh PR replacing #2968, because GitHub could not reopen the old closed PR after the head branch was updated.

What changed

• Add the UniLora tuner implementation, config, layer, and model integration.
• Register PeftType.UNILORA and expose UniLoraConfig / UniLoraModel.
• Add save/load handling for UniLora shared weights.
• Add optional save_indices support for generated index and scale tensors.
• Add package reference documentation.
• Add a quick UniLora fine-tuning example.
• Add a MetaMathQA experiment config.
• Add UniLora coverage in custom model, decoder, encoder-decoder, and dedicated UniLora tests.

Notes for review

This branch is rebased on the latest main and incorporates the remaining review feedback from #2968, including naming, docs, tests, example, MetaMathQA config, deterministic index generation documentation, init_weights, and optional index saving.

Tests

• make style
• make quality
• PYTHONPATH=src python -m pytest tests/test_unilora.py
• PYTHONPATH=src python -m pytest --no-cov tests/test_custom_models.py -k UniLora
• PYTHONPATH=src python -m pytest --no-cov tests/test_decoder_models.py -k UniLora
• PYTHONPATH=src python -m pytest --no-cov tests/test_encoder_decoder_models.py -k UniLora