fix(proxy): preserve image rate-limit lifecycle

[buluw]

Summary

Parse short 429 delays without inflating the grace-retry decision, retry the same account at most once, and preserve only explicit long-lived image quota deadlines.

Closes #3351.
Closes #3353.

Depends on #3362. This is a stacked PR; the new behavior is contained in commit 38a276bf and the diff will narrow after #3362 merges.

Changes

• Parse structured Google retry delays and Retry-After separately from source-specific wait buffers.
• Track one same-account grace retry per account without consuming account-rotation budget.
• Pass retry state through OpenAI Chat, Responses, Gemini, and Images handlers.
• Preserve long deadlines only for explicit structured QUOTA_EXHAUSTED responses on recognized image models.
• Persist and restore qualified image lockouts across quota refresh, account reload, and restart.
• Retain qualified locks through optimistic reset and clear only the matching normalized model bucket.
• Serialize account-file updates so concurrent limit and account-state writes do not discard one another.
• Avoid blocking the async runtime while reselecting after a short lock expires.

Validation

• cargo test --manifest-path src-tauri/Cargo.toml --offline --lib task_
• Result: 11 passed, 0 failed.
• Coverage includes structured 1s/3s parsing, same-account rotation state, explicit 58-hour/72-hour image locks, refresh/reload/restart retention, exact-bucket clearing, and nonblocking short-lock reselection.

Compatibility

• Generic prose, inferred reset times, and text-model limits remain capped at 300 seconds.
• Existing account JSON remains compatible.
• OpenAI Chat, Responses, Gemini, and Images use the new request retry state.
• Claude retains its baseline retry wrapper and parser behavior; same-account grace-retry parity is not claimed.
• No image scheduler or per-account concurrency behavior is included in this PR.