fix: adapt to ChatGPT API v2 metadata format + Turnstile solver support

[abczsl520]

Problem

ChatGPT upstream API changed the metadata format for image generation tool messages:

• Old format: metadata.async_task_type = "image_gen"
• New format: metadata.image_gen_title = "..." (final image) or metadata.async_source = "..." (preview)

This caused ExtractImageToolMsgs to match zero tool messages during poll, making image generation always timeout with no results.

Additionally, the upstream now requires Turnstile challenge solving for the two-step chat-requirements flow (prepare → finalize).

Changes

1. Core fix: metadata compatibility (image.go)

ExtractImageToolMsgs now matches tool messages by any of:

• async_task_type = "image_gen" (old format, backward compatible)
• image_gen_title != "" (new format, final image)
• async_source != "" (new format, preview image)

2. Turnstile solver support

• TurnstileSolver interface: added token parameter to Solve(ctx, dx, token) — the token is the p_value / requirements_token needed for Turnstile bytecode decryption
• New HTTPTurnstileSolver: delegates to an external HTTP microservice (e.g. a Python VM decompiler)
• New config: upstream.turnstile_solver_url — when set, enables the two-step prepare → solve → finalize flow
• Falls back to single-step chat-requirements when solver is not configured or fails

3. Tuning

• Poll interval: 3s → 15s (upstream returns 429 at higher frequencies)
• Consecutive 429 tolerance: 3 → 10 (prevents premature abort)
• PerAttemptTimeout: 6min → 10min (image generation can take 3-5 min with retries)
• PollMaxWait: 300s → 480s

Testing

Tested with a ChatGPT Plus account:

• Turnstile two-step flow: ✅ (token_len ~2900)
• Image generation via /v1/images/generations: ✅ (HTTP 200, 1-2 images returned)
• Image generation via /api/me/playground/image: ✅
• Poll correctly extracts sediment IDs from new metadata format
• Proxy image download via /p/img/ endpoint: ✅