Add frozen model inference engine support for hosting reward models without weight update

[ghShu]

This patch adds support for dedicated reward model inference engines that use frozen_model=True (no weight sync, always active). This enables:

• LLM-as-Judge patterns (RLAIF, Constitutional AI)
• Process Reward Models (verifiers)
• Frozen reward models for scoring/evaluation

Changes:

• Add RewardInferenceConfig and PlacementGenerationEnvConfig to config
• Add pretrained_lora_path option to SkyRLLoraConfig
• Add reward_inference section to ppo_base_config.yaml
• Add get_reward_inference_client() method to BasePPOExp
• Add frozen_model parameter to create_ray_wrapped_inference_engines()
• Pass reward_inference_client to generator