[Bug]: Prompt polishing can spin for minutes with no timeout or cancel action

## Summary

Prompt polishing in resource editors can remain in a loading state for several minutes or appear never to finish. During that time the UI shows only a spinner: it does not identify the model being used, expose request progress, provide a timeout, or let the user cancel.

## Environment

- Cherry Studio version: To be confirmed; 2.0.12 was discussed in the surrounding community conversation, but the reporter did not explicitly confirm their own build
- OS and architecture: To be confirmed
- Provider/model: To be confirmed
- Prompt length: To be confirmed

## Reproduction steps

1. Open a Prompt, Assistant, or Agent resource editor with a non-empty prompt.
2. Click **Polish prompt**.
3. Observe the loading indicator while the generation request is pending.
4. In affected cases, wait several minutes; the action may take about five minutes or appear not to return.

## Actual result

- The polish action can remain pending for minutes or provide no result.
- The user cannot tell which model is handling the request.
- There is no visible timeout, elapsed/progress state, or cancel/retry control while the request is pending.

## Expected result

Prompt polishing should complete within a bounded period or move to a recoverable timeout/error state. The pending UI should identify the selected model and let the user cancel or retry without closing the editor.

## Evidence

- An August 15, 2026 community report observed an approximately five-minute polish operation.
- A separate September 6, 2026 report again described long spinning and, at times, no response.
- Neither report included a log or screenshot for this specific operation.

## Investigation

In the inspected main-line snapshot (`554e42c0a9a40c90f29fa204ac14dbdabfe38806`):

- [`PromptPolishActions`](https://github.com/CherryHQ/cherry-studio/blob/554e42c0a9a40c90f29fa204ac14dbdabfe38806/src/renderer/components/resourceCatalog/dialogs/components/PromptPolishActions.tsx#L81-L157) awaits `fetchGenerate` and keeps the button in a spinner state until the promise settles. It has stale-result guards, but no user cancellation or local timeout.
- [`fetchGenerate`](https://github.com/CherryHQ/cherry-studio/blob/554e42c0a9a40c90f29fa204ac14dbdabfe38806/src/renderer/utils/aiGeneration.ts#L96-L126) resolves an explicitly supplied model or falls back to `readDefaultModel()`. `PromptPolishActions` does not supply a model, so this path uses the default model in that snapshot.

This confirms the unbounded pending-state and model-selection behavior in the inspected snapshot. It does not establish whether the observed delay originates in the provider, network, IPC, or request lifecycle.

## Acceptance criteria

- The polish action has a documented request deadline and exits the loading state when it is exceeded.
- Users can cancel an in-flight polish request and retry without closing the editor.
- The UI identifies the model used for polishing and links or routes to the relevant default-model setting.
- Timeout, provider failure, missing model, and empty response are visibly distinct recoverable states.
- A late response after cancellation or timeout cannot overwrite the user's current prompt.
- Regression tests cover a never-resolving request, a delayed success, cancellation, timeout, and retry.

Start with src/renderer/components/resourceCatalog/dialogs/components/PromptPolishActions.tsx and trace its fetchGenerate call into src/renderer/utils/aiGeneration.ts. Inspect how the pending state and default model are handled, then add regression coverage for never-resolving, delayed, cancelled, timed-out, retried, and late responses. Done means the UI exposes the model, supports recovery, and cannot apply stale results.

Written by the indexing model from the issue text.