feat: improve first-run setup and adaptive benchmarking

[wimi321]

Summary

Makes GoAgent's first launch understandable without turning setup into a hard gate. New users are guided through an optional AI teacher connection and local KataGo readiness, then enter the workbench even when AI setup or benchmarking is deferred.

User Experience

• Adds a versioned, one-time first-run guide in all seven supported UI languages.
• Guides users through AI service URL, access key, model selection, and explicit text/image/tool verification.
• Keeps manual model entry available when /models is unsupported.
• Opens the AI settings page instead of showing technical IPC errors when teacher features are used before configuration.
• Keeps KataGo available when AI setup is skipped.
• Reduces settings to five user-facing pages and moves remote/runtime parameters into advanced analysis settings.
• Removes misleading readiness labels and hides raw technical errors behind expandable details.

KataGo Readiness

• Adds background auto-benchmarking with a 30-second cap and balanced fallback.
• Supports cancelling the current benchmark independently from disabling future automatic benchmarks.
• Makes active game analysis preempt background benchmarking.
• Preserves the previous valid benchmark when a new run is cancelled or fails.
• Adds resumable KataGo model downloads with pause/resume and HTTP Range support.
• Keeps the top analysis speed tied to live search data rather than historical benchmark results.

AI Capability Verification

• Verifies visible text output.
• Verifies actual image recognition using a deterministic color target, not merely image payload acceptance.
• Verifies a correctly formed agent tool call.
• Persists verified, skipped, unconfigured, and needs-attention setup states without using macOS Keychain.

Verification

• pnpm install --frozen-lockfile
• pnpm test (209/209)
• pnpm typecheck
• pnpm build
• pnpm check
• pnpm check:teacher-quality

Manual QA

Validated with a fresh isolated Electron user directory on macOS:

• Seven-language first-run switching without horizontal overflow.
• AI setup can be skipped and KataGo remains ready.
• First-run guide does not return after restart.
• Unconfigured teacher entry opens the AI settings page.
• Settings expose exactly five primary pages.
• Manual benchmark cancellation does not disable future automatic benchmarking.
• Disabling automatic benchmarking persists after restart and prevents background runs.

Issue #23 remains open and now notes the planned Baidu Netdisk mirror after SHA256 verification.