feat: add DynamoModel CRD and controller

[hhzhang16]

Overview:

This MR adds a DynamoModel resource following this enhancement proposal.

Details:

• adds DynamoModel CRD
• adds DynamoModel controller to download models to cache
• updates DGD controller to pull/reference from the downloaded model; waits for model to be "ready" before proceeding

TODO still

• Remove pvc creation and make it optional; without it, it downloads to the node local system (race conditions?)
• Automatic model path argument injection for all backends
• Testing
• Update examples
• add support matrix -- only HF, only PVC RWX

Example CRDs:

# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

# Basic DynamoModel example with HuggingFace public model
apiVersion: nvidia.com/v1alpha1
kind: DynamoModel
metadata:
  name: llama-3-70b-instruct-v1
  namespace: dynamo-cloud
spec:
  # Canonical model name from HuggingFace
  name: meta-llama/Llama-3.3-70B-Instruct
  
  # Version pin (optional but recommended for production)
  # This ensures all deployments use the exact same model artifact
  version: main
  
  # Source URL - supports hf://, s3://, ngc://, or https://
  sourceURL: hf://meta-llama/Llama-3.3-70B-Instruct
  
  # Secret reference for authentication (required for private models)
  secretRef: hf-token-secret
  
  # PVC configuration for model storage
  pvc:
    create: true
    storageClass: standard  # Update with your storage class
    size: 200Gi  # Adjust based on model size
    volumeAccessMode: ReadWriteMany  # Required for multi-replica deployments

---
# DynamoGraphDeployment that uses the model
apiVersion: nvidia.com/v1alpha1
kind: DynamoGraphDeployment
metadata:
  name: vllm-with-model
  namespace: dynamo-cloud
spec:
  # Reference the DynamoModel at top-level (applies to all services)
  modelRef: llama-3-70b-instruct-v1
  backendFramework: vllm
  services:
    VllmWorker:
      replicas: 2
      resources:
        limits:
          nvidia.com/gpu: "2"
      # Model arguments will be auto-injected by the controller
    Frontend:
      replicas: 1

Where should the reviewer start?

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• closes GitHub issue: #xxx