[DOCS] Document Hyperstack disk layout — root disk ~100 GB fixed, use /ephemeral (750 GB) for model caches

[dims]

Summary

Hyperstack instances have a fixed ~100 GB root disk and a 750 GB ephemeral disk at /ephemeral. Neither fact is documented. For model-serving workloads this causes the root disk to fill mid-download with no clear error.

This is the docs companion to #17 (the API/bug side of the same issue).

What needs documenting

On the Hyperstack instance page:

• Root disk: ~100 GiB, fixed — the diskStorage parameter has no effect
• Ephemeral disk: ~750 GiB at /ephemeral, survives reboots, may be wiped on re-provision

For model-serving workloads:

# Wrong — fills the 100 GiB root disk mid-download for any model >50 GB
docker run -v ~/.cache/huggingface:/root/.cache/huggingface ...

# Correct
mkdir -p /ephemeral/huggingface
docker run -v /ephemeral/huggingface:/root/.cache/huggingface ...

Impact

A user downloading an 80 GB model (Llama 70B, Qwen3-Coder-Next-FP8, etc.) to the default HuggingFace cache fills the root disk mid-download. The container crashes with a generic engine error — the actual disk-full warning is buried in container logs. The instance must be deleted and re-created. An hour of GPU time is wasted.

This note should appear on the Hyperstack instance type page and in any quickstart involving large model downloads.