fix(provider): add NVIDIA NIM DeepSeek request compatibility

[fuselayer]

Issue for this PR

Closes #24264

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

DeepSeek V4 models on NVIDIA NIM (deepseek-ai/deepseek-v4-flash, deepseek-ai/deepseek-v4-pro) hang instead of streaming a response. NIM only streams reasoning tokens for these models when the request body includes chat_template_kwargs. Since the AI SDK drops unknown top-level config keys, there's no way to set this from opencode.json, so the field never reaches the wire and the request stalls.

Two changes in transform.ts:

1. In options(), for the nvidia provider on @ai-sdk/openai-compatible, I set chat_template_kwargs: { thinking: true, reasoning_effort: "high" } for deepseek-v4 models. This lands directly on the request body (same pattern as the existing alibaba-cn enable_thinking block right above it), so it isn't stripped by the SDK.

2. In normalizeMessages(), I remap any developer role message to system for the nvidia provider. NIM returns a 500 on the developer role once chat_template_kwargs is present, so without this the first change just turns the hang into a 500.

I confirmed the field names against NVIDIA's own DeepSeek V4 docs (they use chat_template_kwargs with thinking, not enable_thinking, which is a different NIM model family).

How did you verify your code works?

Ran the patched build locally (bun run dev:desktop) against NVIDIA NIM with my own API key.

• Before: deepseek-v4-flash hung until timeout.
• After: it streams normally.
• No more developer-role 500s.
• The only errors I hit afterward were NVIDIA-side quota (ResourceExhausted), which confirms the request is actually reaching the backend now instead of stalling on malformed body.

I couldn't re-verify against NVIDIA-hosted Kimi because that model returns a 404 "function not found for account" for my key right now, so I scoped this PR to DeepSeek only.

Screenshots / recordings

N/A — not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR