Automate local disk management and ai runtime model management

### 🚀 Feature Description and Motivation

In DeepSeek-R1 examples, we notice setting up ReadWriteMany volume is kind of painful for cloud agnostic environments. S3FS based solutions are not very efficient. Most efficiency way is to fully leverage bandwidth to download models from remote object store (0->1) and then load it from disk to GPU memory(cover restarts, redeploy case). However, even most high end machine provides local disk, they are not efficiently managed. We need some local disk management solutions along with runtime cache capabilities (like huggingface cache folder) to streamline this process.

### Use Case

efficiently load model weights from object storage and manage in local disks.

### Proposed Solution

_No response_

No files, tests, or entry points are named. Start by reviewing the DeepSeek-R1 examples and the existing model-loading flow, then examine how S3FS and Hugging Face-style caching relate to local disk handling. Done should include a defined, cloud-agnostic approach for downloading model weights, reusing them across restarts or redeployments, and managing disk capacity.

Written by the indexing model from the issue text.