[Core][Train] Support Torch runtime with TPUs

Description

This can serve as a tracking issue for efforts related to enabling support for PyTorch on TPUs with Ray.

• Inject Torch TPU required vars to RayCluster Pods requesting TPU
  • https://github.com/ai-on-gke/kuberay-tpu-webhook/pull/31
• Fix resource accounting for workloads that require per-device processes
  • https://github.com/ray-project/ray/pull/64058
• Update TorchTrainer to support TPU backend and add validating tests
  • POC: https://github.com/ray-project/ray/pull/63442
  • https://github.com/ray-project/ray/pull/64796
• Implement multi-slice support in Ray Train's TorchTrainer
  • https://github.com/ray-project/ray/pull/66055
• Support MPMD workloads with TorchTPU and Ray
  • Fix TPU_VISIBLE_CHIPS resource bounding for privileged K8s containers on GKE
    • https://github.com/ray-project/ray/pull/65723
  • Support TorchTPU with SubSlicePlacementGroups
    • https://github.com/ray-project/ray/pull/65221
  • Add physical topology ordering to ray.io/tpu-worker-id used by Ray and Torch/vLLM frameworks
    • https://github.com/ray-project/ray/pull/66029

Use case

No response