fix: give vLLM-backed OpenCode providers the generation controls they were silently dropping

[atomantic]

Summary

The seeded OpenCode vLLM providers were the one local OpenAI-compatible runtime whose generation controls did nothing — temperature, top-p, thinking mode and reasoning effort were all silently discarded, and the editor rendered no controls to set in the first place.

• THINKING_STYLE gains a vllm: 'chatTemplate' row. A missing entry is not a missing checkbox: buildAgentGeneration bails on its hasOwn guard and returns null, so temperature, top-p and reasoningEffort went with the toggle. vLLM routes thinking through the chat template exactly as MTPLX and llama.cpp do.
• The client mirrors gain the vllmBacked marker — generationControlsFor (the Generation Defaults block) and isOpencodeLocalProvider (the reasoning-effort ladder). Both were written before vllmBacked existed; the second one was found by the new test table rather than reported.
• apiGenerationOptions gains the same marker. No vLLM api preset ships, but that guard mirrors the client's, so leaving one half narrower re-creates the bug the moment a record is hand-built.
• No provider seed gains a temperature / topP / thinking value. Unset stays unset — an unset control means the container keeps its own chat-template default, which is not the same as being pinned. What was missing is the capability, not a default.
• A new guard walks LOCAL_RUNTIMES and fails if a runtime OpenCode can be pointed at has no THINKING_STYLE entry, so a sixth one cannot land with the same hole.

Docs cleanup folded in from the same #4716 review:

• dflash2.md now points RTX 3090 readers at the container (it linked one way only).
• The bring-up record regains the upstream source links, the batch/ vs single-user/ profile comparison with its context trade-offs, and the cross-links to the sibling research notes — all lost when the doc was rewritten in place.
• The feature doc tells operators to set thinking off + temperature ~0.7 on the provider card, and why both ship unset.

Test plan

• cd server && npm test — 32,746 passing. One pre-existing local failure, routes/imageGen.multipart.test.js (a 10s beforeAll hook timeout, unrelated to this diff); confirmed it fails identically with origin/main's server/ and client/ checked out.
• cd client && npm test — 9,330 passing; npm run lint clean.
• Bypass-probed both new guards: removing the vllm row from THINKING_STYLE turns the LOCAL_RUNTIMES walk red, and removing vllmBacked from the apiGenerationOptions guard turns the new runner test red.

Closes #4765