feat(webui): add preset save & apply for reusable configs

[redXpanda]

Summary

Add a lightweight preset system to the VoxCPM WebUI, mirroring the IndexTTS workflow: save the current reference audio + all generation parameters as a named preset and reload it later. Also localizes preset toast messages so they match the gr.I18n-driven UI.

What's included

• presets.py — file-based preset store (presets/<name>/preset.json + copied reference audio), with name sanitization and path resolution.
• app.py — inline preset UI (name + save / dropdown + apply / delete / refresh) under the reference audio, plus i18n keys (en + zh-CN).
• .gitignore — ignore presets/ (user data).

Notes

• A preset stores the reference audio, cloning-mode toggle and transcript, the control instruction, the target text, and the advanced generation parameters.
• Toast messages (save / apply / delete) were previously hardcoded in Chinese. Since gr.Info/gr.Warning reject I18nData (they require a plain str), the language is resolved from the request's Accept-Language header and the message is looked up in a runtime dict.

Test plan

• Save a preset from the WebUI and confirm presets/<name>/preset.json + the copied reference audio are created.
• Apply the saved preset and verify all fields (reference audio, transcript, instruction, target text, advanced params) are restored.
• Delete / refresh the preset dropdown.
• Switch browser language (en / zh-CN) and confirm toast messages are localized correctly.

[a710128]

Nice work on this — the preset system is clean and well-documented, and I really like that the save inputs / apply outputs share a single source-of-truth ordering so they can't drift apart. I pulled it down and gave it a read + a quick functional roundtrip; everything lines up nicely. 🙌

One small suggestion: would you consider adding a unit test for presets.py?

The repo already has a tests/ directory with a pytest setup (pytest + pytest-cov in pyproject.toml), and presets.py is a great fit for testing since it's pure logic with no heavy deps — no torch, no gradio, no model loading — so the tests would run fast and GPU-free in any environment.

A few high-value cases worth locking down:

• safe_preset_name edge cases (empty / whitespace-only names, purely-special-char names, / \ : .., and CJK input) — this is effectively the path-traversal guard, so it's the most valuable thing to pin down.
• save → load roundtrip: confirm all params are preserved and the audio file suffix is kept (e.g. a .flac reference stays .flac).
• reference_audio=None → stored as "".
• delete_preset returns the correct bool for existing vs. missing presets.
• load_preset returns None for a non-existent preset.

Using tmp_path (or monkeypatching get_presets_dir) keeps the tests isolated from the real presets/ directory.

I'd skip testing the app.py handlers themselves — they're thin gradio glue wrapping gr.update() / gr.Warning / gr.Request, so mocking the runtime there isn't worth the cost.

Totally non-blocking from my side — just thought it'd be a nice fit given the existing test conventions. Thanks for the contribution!