[original4422]

Summary

Explain the existing server request interruption behavior, as requested in #897:

 Server options
+ Streaming interruptions: default True, queued mode with False, and serialization limits

Add CLI and top-level JSON examples for letting the current stream finish before waiting requests run.
Clarify that requests still share a model proxy lock, and client disconnections or other cancellation can still end a stream.
This documents existing behavior; it does not change defaults or implement parallel decoding.

Evidence

• Before: the server page had a generated settings reference but no explanation of stream interruption or queueing.
  After: the new section and its CLI/JSON examples render in the actual MkDocs page, with settings anchors preserved.
• Original-package CLI parsing and Settings, ServerSettings and ConfigFileSettings checks confirm the documented True default and explicit false values.
• python -m mkdocs build --clean and make lint pass. The baseline's 8 focused original tests also pass.
• Existing MkDocs deprecation/link/Griffe warnings remain. Full model-download tests and inference/GPU/concurrency performance tests were not run for this documentation change.

AI-assisted documentation change.

Merge Danger

Door: two-way

Blast Radius: documentation

One server documentation page changes; runtime code and dependencies stay unchanged.