Move LoRA config to TrainableModel

[corbt]

Summary

• Add top-level TrainableModel.lora_config and public art.LoRAConfig with normalized LoRA field names.
• Replace _internal_config.peft_args with normalized internal lora_config, and raise a descriptive migration error for legacy peft_args / art.dev.PeftArgs(...) usage.
• Wire normalized LoRA config through the Megatron trainer, keeping PEFT-shaped names only at PEFT/Unsloth adapter boundaries, and update repo examples/tests.

Tests

• uv run ruff check src/art/dev/get_model_config.py src/art/dev/model.py src/art/model.py src/art/megatron/lora.py src/art/megatron/service.py src/art/megatron/train.py src/art/unsloth/service.py tests/unit/test_dedicated_config.py examples/hn_title_generator/train.py dev/yes-no-maybe.py
• uv run pytest tests/unit/test_dedicated_config.py -q
• uv run ty check src/art/dev/get_model_config.py src/art/dev/model.py src/art/model.py src/art/megatron/lora.py src/art/megatron/service.py src/art/megatron/train.py src/art/unsloth/service.py tests/unit/test_dedicated_config.py
• commit hooks: ruff, ruff format, ty check src tests, uv lock --check