fix(provider): patch empty Gemini replay messages for Vertex AI

[ghchinoy]

Issue for this PR

Closes #47863

Type of change

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

What does this PR do?

PR #30463 patched @ai-sdk/google@3.0.73 to pop empty model turns (if (contents[contents.length - 1].parts.length === 0) contents.pop()). However, @ai-sdk/google-vertex depends on @ai-sdk/google@3.0.108 in bun.lock, which was not covered by the 3.0.73 patch rule. As a result, Gemini on Vertex AI (google-vertex) still emitted { role: "model", parts: [] } when an assistant turn contained only empty reasoning parts (text: ""), resulting in HTTP 400 (Unable to submit request because it must include at least one parts field).

This PR:

1. Adds patches/@ai-sdk%2Fgoogle@3.0.108.patch to apply the same contents.pop() guard to @ai-sdk/google@3.0.108 used by @ai-sdk/google-vertex.
2. Adds a guard in MessageV2.toModelMessagesEffect (packages/opencode/src/session/message-v2.ts) to skip zero-length reasoning parts when replaying history.
3. Adds a unit test in packages/opencode/test/session/llm.test.ts covering google-vertex payload construction with empty reasoning parts.

How did you verify your code works?

• Ran bun test test/session/llm.test.ts -t "sends Google": both google and google-vertex tests pass.
• Verified typecheck with bun turbo typecheck (all 30 package tasks passed).

Screenshots / recordings

N/A (backend/protocol fix)

Checklist

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