docs(sdk): describe typed provider boundaries and Responses streaming

[vedjoshi1]

• I have read and reviewed the documentation changes to the best of my ability.
• If the change is significant, I have run the documentation site locally and confirmed it renders as expected.

Summary of changes

• Describe private typed provider boundaries and optional chat-template tokenizer
  loading/fallback behavior in the SDK LLM architecture page.
• Document Responses streaming, sync/async callbacks, required-stream endpoints,
  generic iterator support, and yielded-versus-wrapper completion precedence.

Companion implementation: OpenHands/software-agent-sdk#5504
Tracking issue: OpenHands/software-agent-sdk#4976
Completion-selection behavior follows alanhuangyoo's SDK PR #4772.

Validation / limitations

• Fresh local docs checks after the example correction: uv run --with pytest --with requests --with pyyaml pytest -q tests/ produced 35 passed and 2 errors. The errors were pricing
  fixtures fetching a removed upstream file (HTTP 404); these pages do not change
  pricing documentation or tests.
• Companion SDK validation includes real local HTTP/SSE parsing and packaged
  server startup; detailed results are in the SDK PR.
• The updated Responses example was syntax-checked and executed against the
  installed SDK with a recording stand-in for the request, confirming imports,
  streaming arguments, and system-before-user message order. No live model call
  was made. Mintlify rendering has not been verified locally.
• The human contributor reviewed the documentation changes. Both linked PRs
  are ready for maintainer review; the SDK branch has been rebased and validated.

[neubig]

Approving on behalf of @neubig.

Reviewed the diff and verified every documented claim against the merged SDK code (main @ 69eb315, the squash commit of #5504):

• Example is correct and runnable. Ran the new responses() snippet verbatim against a real model (gpt-5.5 via the Responses endpoint, stream=True, on_token): it streamed output and returned an LLMResponse. ModelResponseStream is genuinely re-exported from openhands.sdk.llm.streaming, so the import resolves.
• Removed warning was accurate to remove. responses()/aresponses() do accept stream=True; the old 'chat completion endpoint only' caveat was stale. No other stale limitation text remains on the page.
• Behavioral prose matches implementation: non-streaming fallback when on_token is None (llm.py:1790, 2048); requires_streaming/is_subscription draining without a callback (llm.py:961); LLMNoResponseError on a stream with no completed response (llm.py:1337, _response_stream.py:57); wrapper completed_response precedence and generic sync/async iterable support (_response_stream.py); optional Transformers tokenizer with LiteLLM fallback (_tokenizer.py).
• Docs tests reproduce exactly as described: 35 passed, 2 errors. I confirmed the 2 errors are identical on unmodified main (pricing fixtures 404 on a removed upstream file) — pre-existing and unrelated to this PR.
• Rendering/structure: code fences balanced, both pages present in docs.json nav, the new /sdk/guides/llm-streaming link and #ready-to-run-example anchor both resolve. The new code block carries no file-reference metadata, so the daily sync-docs-code-blocks job will not overwrite it.

Note: the broken-links workflow is still gated as action_required (fork PR); I could not approve it with the available token, so link validation above was done manually.

This review was produced by an AI agent (OpenHands) on behalf of @neubig.