Add GPU / TPU / accelerator targeting and resource injection

[tartkri-bot]

Problem

The predictor topology reads comp.Resources but does not provide any GPU (NVIDIA GPU / TPU / AWS Inferentia / Trainium) resource helpers, and does not auto-inject or validate GPU-specific requirements. KServe users frequently need:

1. GPU resource requests: Explicitly requesting nvidia.com/gpu: "1" or amd.com/gpu in the container resources.
2. Auto-GPU detection: Based on model format or selected ServingRuntime, optionally suggest/add GPU resources automatically.
3. TPU support (Google): Selector for cloud.google.com/gke-accelerator and toleration for cloud.google.com/gke-tpu.
4. AWS Inferentia / Trainium: Resources like aws.amazon.com/neuron and node selectors for Neuron nodes.
5. GPU-specific environment variables: Auto-inject CUDA_VISIBLE_DEVICES, NVIDIA_VISIBLE_DEVICES, or runtime-specific vars when GPU is requested.

Currently, users must manually set comp.Resources with the exact GPU resource name and quantity, which is Kubernetes-specific and error-prone.

Suggested approach

Add a gpu or accelerator block to ModelServerCustomSpec (and LLM parameters):

• type — "nvidia", "tpu", "inferentia", "trainium", "amd", etc.
• count — number of accelerators.
• autoInjectEnv — boolean, defaults to true. When enabled, inject common env vars for the selected accelerator type.

Alternatively, keep it simpler: add nvidiaGPU, tpu, neuron as optional typed fields with count + autoEnv.

Acceptance criteria

• GPU resources are correctly injected into the serving container's resource requests/limits.
• Node selectors / tolerations for the selected accelerator type are auto-applied where applicable.
• Accelerator-specific env vars are injected when requested.
• Multi-GPU scenarios (tensor parallelism across GPUs on the same node) are validated.
• Chainsaw tests verify GPU scheduling on clusters with GPU nodes (or mock with node labels for CI).