Edge runtime improvements

We made the edge runtime lighter and more observable.

Performance

• Cold start time reduced by 23% on Xavier-class devices.
• Streaming now defaults to chunked responses for chat completions.

Reliability

• Added backpressure handling on low-bandwidth networks.
• Improved retry heuristics for transient storage failures.

Tooling

• unieai deploy --edge now surfaces GPU temperature and throttling warnings inline.