lazy-load model handlers to decouple megatron deps from tinker path

[hardik-vala]

Summary

• from art.tinker import TinkerBackend; await model.register(backend) previously failed with ModuleNotFoundError: No module named 'megatron' on environments without megatron-core installed (e.g. macOS, where the [megatron] extra can't install due to transformer-engine / apex / deep-ep).
• the chain was tinker._get_service → art.dev.get_model_config → art.megatron.model_support.registry, whose top-level handler imports pulled handlers/qwen3_5.py and its top-level from megatron.core....
• introduces HandlerMeta class in spec.py as the single source of truth for each registered handler's key, native_vllm_lora_status, and lazy-load address (module, attr). registry.py builds every ModelSupportSpec from these metas and resolves handlers lazily via importlib.import_module on the first call to get_model_support_handler*. The handler classes also read their key / native_vllm_lora_status from the same metas so the values can't drift.
• net effect: import art.tinker, from art.dev.get_model_config import get_model_config, and TinkerBackend._get_service(model) all succeed without megatron-core. Megatron-equipped paths are unchanged — same handler instances, just imported on first lookup instead of at registry load.

Test plan

• On a venv with pip install -e '.[tinker]' and no megatron-core:
  • python -c "import art.tinker.backend" succeeds
  • get_model_config(base_model='Qwen/Qwen3-4B-Instruct-2507', ...) returns the expected
    peft_args["target_modules"]
  • default_target_modules_for_model('Qwen/Qwen3-30B-A3B-Instruct-2507') returns [..., 'experts']
  • await TinkerBackend()._get_service(model) constructs TinkerService without importing megatron
• uv run prek run --all-files clean