feat: Allow to set gpu and worker scheduling params for ray remote tasks

[ntkathole]

What this PR does / why we need it:

This PR adds first-class GPU support and a general worker_task_options escape
hatch for any Ray .options() scheduling parameter, wired consistently
through all execution modes (local, remote, KubeRay).

Example config

batch_engine:
    type: ray.engine
    num_gpus: 1
    gpu_batch_format: numpy
    worker_task_options:
        num_cpus: 4
        memory: 8589934592        # 8 GB
        accelerator_type: "A100"  # pin to A100 on mixed GPU pools (KubeRay)
        max_retries: 5
        runtime_env:
            pip: ["cudf-cu12==24.10.0"]
            env_vars: {CUDA_VISIBLE_DEVICES: "0"}

---