fix(opencode): bump togetherai so streams report usage

[rekram1-node]

Fixes #47716.

Summary

V1 records 0 tokens and $0 cost for Together models because @ai-sdk/togetherai@2.0.41 never sends stream_options.include_usage. Together only returns usage when that flag is set.

This bumps the bundled package to 2.0.68 in packages/opencode and packages/core. That release hardcodes includeUsage: true on chat and completion models. Passing the flag ourselves does not work: createTogetherAI only reads apiKey, baseURL, headers, and fetch.

We are already on the AI SDK v6 line (@ai-sdk/togetherai 2.x, @ai-sdk/provider 3.x). 3.0.45 is the next major and pulls @ai-sdk/provider 4, so this stays on 2.0.68, the patch that added the flag.

Review

• Public TogetherAIProviderSettings and languageModel / chatModel shapes are unchanged from 2.0.41.
• Nested @ai-sdk/openai-compatible moves from 2.0.37 to 2.0.62. Chat request fields OpenCode relies on (max_tokens, tools, reasoning_effort) are unchanged. Tool-call streaming is stricter in a compatible way, and SSE errors now forward the error object, which matches the local patch already applied to direct @ai-sdk/openai-compatible@2.0.41.
• Together nests @ai-sdk/provider@3.0.14 while the app uses 3.0.16. LanguageModelV3 is unchanged across that range, and this path does not instanceof it.
• Lockfile blast radius is the togetherai alias plus its nested openai-compatible@2.0.62, provider@3.0.14, and provider-utils@4.0.40. Direct provider pins are unchanged. An unused vitest coverage subgraph dropped out of the lockfile.

Test plan

• Together streaming request body includes stream_options.include_usage: true
• Assistant message usage and cost are non-zero for a Together chat model
• A Together tool-call turn still completes and records usage