Bug: Voice mode cannot be enabled - Failed to fetch model catalog (catalog unreachable on corporate VPN)

## Bug: Cannot enable voice mode — "Failed to fetch model catalog" / "Couldn''t load voice model catalog"

### Summary
Activating voice mode with `/voice` fails because the CLI cannot fetch the voice (STT) model catalog. Voice mode can''t be enabled at all, since the catalog must be reachable before any speech-to-text model can be downloaded.

### What I see
Two dialogs appear when trying to enable voice mode:

**1) Voice models picker:**
```
Voice models

No STT models available.
Your CLI may be out of date. Run /update to update, or download from:
https://github.com/github/copilot-cli/releases

Failed to fetch model catalog. Press r to retry.
```

**2) Blocking dialog:**
```
Couldn''t load voice model catalog
The voice model catalog couldn''t be fetched.

Voice mode can''t be enabled until the catalog is reachable. Check your network connection and retry.

> Retry
Cancel
```

Pressing `r` / **Retry** does not resolve it — the catalog still fails to load and no STT models become available, so voice mode never enables.

### Steps to reproduce
1. Run the Copilot CLI.
2. Invoke `/voice` to enable voice mode (dictation transcription via Foundry Local).
3. The CLI attempts to fetch the voice model catalog and fails with the errors above.

### Environment
- OS: Windows 11 (arm64)
- Copilot CLI version: 1.0.57
- Network: corporate environment. The machine was connected to a corporate VPN (Azure VPN). The failure occurs in this restricted-network context, which strongly suggests the model-catalog / model-download endpoints are being blocked by the VPN/proxy/firewall.

### Expected behavior
- Voice mode should be able to fetch the model catalog and download STT models in corporate/proxy/VPN environments, OR
- The error should clearly identify the exact endpoint(s) it''s trying to reach (host/URL) and respect system/HTTP(S) proxy settings, so the URLs can be allow-listed.

### Impact
Voice mode is completely unusable in restricted-network/VPN environments because the catalog fetch is a hard prerequisite and there is no offline / manual model-install fallback surfaced in the UI.

### Suggested improvements
- Surface the exact catalog/model URLs in the error so they can be allow-listed.
- Honor `HTTPS_PROXY` / system proxy and custom CA settings for the catalog + model download.
- Provide an offline / manual model-install path when the catalog is unreachable.

### Screenshots

Start with the /voice flow and reproduce on Windows 11 arm64 while connected to the corporate Azure VPN. Trace the Foundry Local voice model catalog request and model-download path, recording the endpoint and proxy or custom-CA behavior. Done means voice mode can load STT models in the reported restricted network or the UI exposes actionable endpoint and proxy guidance.

由索引模型根據 Issue 內容生成。