feat: LoRA self-reference RL — use policy trainer as KL reference

[mayinghan]

When training with LoRA, the base model (without adapter) can serve as the KL divergence reference model. This eliminates the need for a separate FORWARD_ONLY reference trainer job, halving reference GPU cost.

Changes:

• Add BaseReferenceClient: lightweight forward-only wrapper for base model logprobs, sharing the policy client's service session
• Add ReconnectableClient.create_base_reference(): creates a base-only model handle (base-) on the same LORA_TRAINER, with LoRA adapters disabled for forward passes
• Store _svc on ReconnectableClient so the service session can be reused for both policy and reference model handles
• New cookbook example: training/examples/rl/lora_self_ref/ demonstrating GRPO with LoRA where no separate reference training shape is needed

Depends on SDK changes:

• tinker CreateModelRequest.base_only field
• FiretitanServiceClient.create_base_training_client() method